PDF

TECHNISCHE UNIVERSITÄT MÜNCHEN
Lehrstuhl für Angewandte Mechanik
Design and Realization of a

Humanoid Robot for
Fast and Autonomous Bipedal Locomotion
Entwurf und Realisierung eines Humanoiden Roboters für Schnelles
und Autonomes Laufen
Dipl.-Ing. Univ. Sebastian Lohmeier
Vollständiger Abdruck der von der Fakultät für Maschinenwesen der

Technischen Universität München zur Erlangung des
akademischen Grades eines
Doktor-Ingenieurs (Dr.-Ing.)
genehmigten Dissertation.
Vorsitzender: Univ.-Prof. Dr.-Ing. Udo Lindemann
Prüfer der Dissertation:
1. Univ.-Prof. Dr.-Ing. habil. Heinz Ulbrich
2. Univ.-Prof. Dr.-Ing. Horst Baier
Die Dissertation wurde am 2. Juni 2010 bei der Technischen Universität München
eingereicht und durch die Fakultät für Maschinenwesen am 21. Oktober 2010
angenommen.
Colophon
The original source for this thesis was edited in GNU Emacs and aucTEX, typeset using
pdfLATEX in an automated process using GNU make, and output as PDF. The document
was compiled with the LATEX 2ε class AMdiss (based on the KOMA- Script class scrreprt).
AMdiss is part of the AMclasses bundle that was developed by the author for writing
term papers, Diploma theses and dissertations at the Institute of Applied Mechanics,
Technische Universität München. Photographs and CAD screenshots were processed and
enhanced with THE GIMP. Most vector graphics were drawn with CorelDraw X3, exported
as Encapsulated PostScript, and edited with psfrag to obtain high-quality labeling. Some
smaller and text-heavy graphics (flowcharts, etc.), as well as diagrams were created using
PSTricks. The plot raw data were preprocessed with Matlab. In order to use the PostScript-
based LATEX packages with pdfLATEX, a toolchain based on pst-pdf and Ghostscript was
used. The bibliography was prepared with biblatex.
The fonts in this thesis are the light versions of the Kp-Fonts by the Johannes Kepler
project.
© 2010 by Sebastian Lohmeier. All rights reserved
Author contact: [email protected]
This thesis is also available as a printed book published by

Verlag Dr. Hut München, Germany (ISBN 978-3-86853-734-5).
To my parents
The danger of the past was that men became
slaves. The danger of the future is that men may
become robots. True enough, robots do not rebel.
But given man’s nature, robots cannot live and
remain sane, they become ‘Golems,’ they will
destroy their world and themselves because they
cannot stand any longer the boredom of a
meaningless life.
—Erich Fromm, The Sane Society
Abstract
This thesis covers the mechatronic conception, design and realization of a hu-
manoid walking robot. Weighing approximately 60 kg, the robot stands 180 cm
tall and has 25 actuated degrees of freedom in a redundant configuration. The
mechanical structure is characterized by the consistent lightweight design with
high effective stiffness. The lightweight servo actuators combine high-dynamic
brushless servo motors with precision gearings and sensors into a compact pack-
age. Moreover, the resultant inertia of the legs is minimized by a sophisticated
design of the structure and drive mechanisms, resulting in a superior acceleration
behavior. The sensor layout comprises angular sensors for direct measurement of
the joint angles, a high-precision inertial measurement system and force/torque
sensors in the feet. Using the trajectory generation and control system which is
not in the scope of this work, the current speed of locomotion is 3.34 km/h.
Keywords: humanoid robots, anthropomorphic robots, legged locomotion, mecha-

tronics, robot design, sensors
Zusammenfassung
Diese Arbeit beschreibt die mechatronische Konzeption, Konstruktion und Re-
alisierung eines humanoiden Laufroboters. Der Roboter ist 180 cm groß, wiegt
zirka 60 kg und verfügt über 25 angetriebene Freiheitsgrade in einer redundanten
Konfiguration. Die mechanische Struktur ist durch den extremen Leichtbau mit
hoher effektiver Steifigkeit charakterisiert. Die hochdynamischen Gelenkantriebe
kombinieren bürstenlose Servomotoren, Untersetzungsgetriebe und Sensorik zu
kompakten Einheiten. Daneben werden die Massenträgheitsmomente der Beine
durch konstruktive Maßnahmen minimiert. Daraus resultiert ein überragendes
Beschleunigungsverhalten. Neben Winkelsensoren zur direkten Messung der Ge-
lenkwinkel verfügt der Roboter über ein hochgenaues inertiales Messsystem und
Kraftsensoren in den Füßen. Mit dem Bahnplanungs- und Regelungssystem, das
nicht Bestandteil dieser Arbeit ist, konnte bisher eine Maximalgeschwindigkeit
von 3,34 km/h erreicht werden.
Stichworte: humanoide Roboter, anthropomorphe Roboter, zweibeiniges Laufen,

Mechatronik, Konstruktion, Sensorik
vii
Acknowledgments
This thesis summarizes my research carried out at the Institute of Applied Me-
chanics, Technische Universität München. Developing the humanoid robot Lola
almost from scratch has been a unique opportunity and certainly one of the biggest
challenges in my life so far. Many people have contributed to this thesis and I
would like to use the opportunity to express my gratitude to some of them.
First and foremost I want to express my deep and sincere gratitude to my advisor
Professor Heinz Ulbrich for the excellent research environment and giving enough
freedom for own creativity. His patience and irrefutable confidence in my skills
were beneficial for the success of this research.
I also wish to express my warm and sincere thanks to Professor Friedrich
Pfeiffer, the former head of the institute, for a lot of good advice and ideas and
many helpful discussions.
I warmly thank our academic director, Dr. Thomas Thümmel, for his support
and continuing commitment to the robotics group. Thomas managed our project
resources and largely kept us free from most administrative and formal burdens.
Starting off with only two doctoral candidates, the Lola project was divided
simply into “Software” and “Hardware.” I am particularly indebted to Thomas
Buschmann, my co-worker and good friend, who took over the “Software” part.
His theoretically accurate and thorough works on the robot simulation and con-
troller were one of the foundations for the success of the project. To me, working
together with Thomas was a stroke of luck as our ways of thinking and working
were quite similar and, where not, complemented each other.
I really need to thank Markus Schwienbacher and Valerio Favot who joined the
team in a later project phase. In his Diploma thesis Markus developed, assembled
and calibrated the force/torque sensors with enthusiasm and great attention to
detail. I appreciate his efforts in resolving Lola’s teething problems. Markus, thank
you so much for the fruitful discussions and competent suggestions! Valerio tested
the PCBs for the local controllers and wrote all the code for them. Also, special
thanks are due to Mathias Bachmayer, who developed the DSP boards for Lola’s
decentralized controllers.
The official presentation of Lola at the HANNOVER MESSE 2010 was the grand
finale before turning in this dissertation. Regarding our trade fair presence, I
thank Alexander Ewald for taking over all the organizational stuff. And my
thanks go to the developers of Lola’s visual perception system, Dr. Felix von
Hundelshausen and Gerhard Rohe from the Institute of Autonomous Systems
Technology, University of the Bundeswehr, München, for the pleasant cooperation.
During my first year at the institute I have had the pleasure to work with Dr.
Klaus Löffler, the developer of Johnnie’s balance controller. Klaus shared the office
with me and has always been a competent and excellent advisor as I took my first
tentative steps into the field of humanoid robotics. Thank you for the collaboration
ix
x Acknowledgments
during the countless public performances of Johnnie in 2003/04. The designer of

Johnnie’s mechatronic system, Dr. Michael Gienger, had already left the institute
before I started. Nevertheless, he was a competent advisor and always there to
discuss about biped robots. Klaus and Michael, your results and experience were
more than helpful and contributed substantially to this research!
I warmly need to thank Dr. Andreas Zagler for getting me excited about robotics
and for encouraging me to take up this position.
The countless customized mechanical and electronic components would not
have been possible without the enthusiasm and willingness of the institute’s me-
chanical and electronic workshops. My sincere thanks are due to Wilhelm Miller,
Walter Wöß, Simon Gerer, Philipp Schneider and Tobias Schmid. I very much
appreciated their amazing skills to machine even the most complex parts with
unparalleled precision. Until his retirement, Walter was a great help during the as-
sembly of Lola. Wilhelm Miller, the former manager of the mechanical workshop,
deserves a special mention. He fostered the manufacturing with optimism and
never ending perseverance. He was always there to talk about my ideas and solve
any unsolvable problems concerning precision mechanics. I have fond memories
of numerous discussions—be it technical or non-technical. Sadly, he was not given
the time to see Lola walking: he passed away during the first weeks of assembly. I
am also greatly indebted to our electronics engineer, Georg Mayr, for numerous
helpful discussions and for developing most of the electronics hardware. His
experience and openness for my ideas were beneficial.
For this dissertation I would like to thank my proof-readers: Michael, Markus,
Thomas Buschmann, Dr. Martin Schwaiger and Valerio for their time, interest and
many helpful comments.
I would also like to thank the other two members of my dissertation commit-
tee, the second advisor Professor Horst Baier and the chairman Professor Udo
Lindemann, for their time and insightful questions.
I gratefully acknowledge the funding sources that made my dissertation possible.
For most of the time I was funded by the German Research Foundation (DFG)
within the project-cluster „Natur und Technik Intelligenten Laufens“(PAK-146,
grant UL 105/29).
Putting a lot of effort into sophisticated hardware design is very expensive and
makes it difficult for a small university group like ours to compete with larger
research groups and companies with considerable financial support. I like to
thank all companies that supplied us with customized components to very good
conditions. I particularly thank Arne Laugwitz and his colleagues at ACTech
GmbH, Freiberg, who manufactured the investment castings. I appreciated their
collaboration and enthusiasm to pushing the limits of current casting and machin-
ing technologies. Lola’s leg architecture would not have been possible without
their generous support.
Finally, in deep gratitude I want to thank my parents, Gertraud and Hans, for
their never-ending love, encouragement and support in all my efforts.
And most of all for my loving, encouraging and patient wife Veronika who
reminded me that there is a life besides research. Her faithful support during the
final stages of this dissertation is so appreciated.
Contents
Abstract vii
Zusammenfassung vii
Acknowledgments ix
1 Introduction 1
1.1 Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.1 Fully Actuated Biped Walking Robots . . . . . . . . . . . . . . . . . 3
1.1.2 Balance Control of Fully Actuated Biped Robots . . . . . . . . . . . 10
1.1.3 Passive Dynamic Walkers . . . . . . . . . . . . . . . . . . . . . . . . 11
1.1.4 Other Relevant Developments . . . . . . . . . . . . . . . . . . . . . . 13
1.2 Contributions of this Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2 Mechatronic Design Concept 17

2.1 Systematic Approach to Managing Complexity . . . . . . . . . . . . . . . . 18
2.1.1 Methodology in the Development Process . . . . . . . . . . . . . . . 18
2.1.2 Project Planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Kinematic Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.2.1 Geometric Dimensions of the Links . . . . . . . . . . . . . . . . . . . 20
2.2.2 Topology of the Kinematic Chain . . . . . . . . . . . . . . . . . . . . 20
2.2.3 Camera Head . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.2.4 Arm Joints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2.5 Pelvis Joint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.2.6 Toe Joint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3 Foot Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3.1 Biomechanics of the Human Feet . . . . . . . . . . . . . . . . . . . . 27
2.3.2 Design Concept of the Robotic Feet . . . . . . . . . . . . . . . . . . . 28
2.4 General Requirements on a High-Dynamic Leg Architecture . . . . . . . . 30
2.4.1 High Effective Stiffness . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4.2 Total Center of Mass Height . . . . . . . . . . . . . . . . . . . . . . . 35
2.4.3 Low-inertia Design of the Leg Links . . . . . . . . . . . . . . . . . . 36
2.5 Joint Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.5.1 Actuators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.5.2 Drive Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.6 Kinematic Structure of the Leg Joints . . . . . . . . . . . . . . . . . . . . . . 50
2.6.1 Hip Joint Actuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
2.6.2 Knee Joint Actuation . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
2.6.3 Ankle Joint Actuation . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.6.4 Toe Joint Actuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.7 Sensor Layout . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.7.1 Upper Body Position and Orientation . . . . . . . . . . . . . . . . . . 64
2.7.2 Joint Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
xi
xii Contents
2.7.3 External Forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

2.7.4 Contact State of the Feet . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.8 Computer Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.8.1 Design Concept . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
2.8.2 Central Control Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
2.8.3 Decentral Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
2.8.4 Communication System . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.9 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3 Mechanical Design 77
3.1 Design Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
3.2 General Design Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
3.3 Design Calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.3.1 Calculation of Fundamental Design Data . . . . . . . . . . . . . . . 81
3.3.2 Dimensioning of the Drive Mechanisms . . . . . . . . . . . . . . . . 82
3.3.3 Dimensioning of Structural Components . . . . . . . . . . . . . . . . 91
3.4 Modular Joint Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.4.1 Motivation for Modularity and its Limitations . . . . . . . . . . . . . 96
3.4.2 Revolute Joints based on Harmonic Drive Gears . . . . . . . . . . . . 97
3.5 The Stereo Robotic Head . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
3.6 The Torso and Pelvis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
3.7 The Arms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.8 The Legs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.8.1 The Hip Joint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.8.2 The Knee Joint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
3.8.3 The Ankle Joint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
3.9 The Feet . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
3.10 The Link Segments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
3.11 Parameter Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3.11.1 Kinematic Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . 124
3.11.2 Link Inertial Parameter Estimation . . . . . . . . . . . . . . . . . . . 125
3.12 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
3.12.1 Mass Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
3.12.2 Analysis of Dynamic Performance . . . . . . . . . . . . . . . . . . . 127
3.13 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
4 Sensor System 133

4.1 Joint Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
4.2 Force/Torque Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.2.1 Optimized Shear Beam Geometry for High Sensitivity . . . . . . . . 136
4.2.2 Mechanical Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
4.2.3 Sensor Electronics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.2.4 Force Sensor Calibration . . . . . . . . . . . . . . . . . . . . . . . . . 148
4.3 Contact Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
4.4 Inertial Measurement System . . . . . . . . . . . . . . . . . . . . . . . . . . 153
4.5 Chapter Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
5 Electronics Architecture 159

5.1 Central Control Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
5.2 Local Controllers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Contents xiii
5.3 Communication System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

5.4 Security System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
6 Lessons Learned 167

6.1 Conceptual Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
6.2 Mechanical System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
6.3 Sensor System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
6.4 Electronics Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
7 Conclusions and Future Work 171
A Anatomic and Biomechanical Fundamentals of Human Locomotion 175
B Main Dimensions 177
C Technical Data of the Joint Actuators 179

C.1 Mechanical . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
C.2 Electrical . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
D Load Profiles 181
Bibliography 189
1 Introduction
In the Fifties, it was predicted that in 5 years robots
would be everywhere. In the Sixties, it was predicted
that in 10 years robots would be everywhere. In the
Seventies, it was predicted that in 20 years robots
would be everwhere. In the Eighties, it was predicted
that in 40 years robots would be everywhere . . .
—Marvin Minsky
No matter whether one believes in God’s creation or theory of evolution: everyone

will agree that nature is fascinating and in a sense “perfect.” We are surrounded
by highly developed creatures that are fully adapted to their environment and
able to react to any kind of environmental influence. What we take for granted
by nature is extremely difficult to realize in technical systems. For example, the
complexity of everyday tasks like “going straight on”—“put one foot in front of
the other . . . just walk!” is anything but simple! Accordingly, interaction between
the “walking tool” leg and the body is highly complex: the body must be kept
in balance at any time; environmental influence such as wind or the nature of
the walking surface must be taken into consideration; the goal is to be focussed,
potential obstacles must be identified and alternative paths must be assessed.
The term “robot” (Czech for “worker”) in its today’s meaning was invented
in 1920 by the Czech author Karel Čapek He wrote satires on technological
progress. In his fictional play “R.U.R.—Rossum’s Universal Robots,” the first
robots were shaped like humans and created for factory work [72]. Eighty-nine
years later, human-shaped robots are still rare, although the number of research
groups worldwide has increased exponentially during the recent years. Some
Japanese companies have been conducting research on humanoid robots for a long
term. They began delivering them in 2004 as human companions for performing
typical everyday tasks. But, according to a study of the International Federation
of Robotics [80], only about 100 are in use, most of them in research laboratories.
Recent studies notwithstanding acknowledge that humanoid service robots might
be an emerging application area in the near future [80, 150]. The anthropomorphic
structure is one reason why such devices are ideal general-purpose robots to
assisting humans at everyday tasks: the human-like structure allows them to
easily operate in and interact with human-built environments. Other scenarios
include medical engineering and prosthetics, the entertainment industry and
academic education and research. For all such tasks in production, office or
home environments robust, fast and agile locomotion is a basic skill. Moreover,
research on walking robots can be seen as a driver of technology in the field of
mechatronics, pushing lightweight drive systems and components, sensors, and
automatic control engineering.
Despite extensive research in the fields of biomechanics and movement sciences,
1
2 1 Introduction
the question how humans solve the highly complex problem of three-dimensional
bipedal locomotion with unparalleled robustness is not well understood to this
day. Although the biomechanics of human walking and running has been investi-
gated quite well [129, 142], the mechanisms of balance control are not very well
understood. For a long time posture control during standing and walking has
been considered an automatic or reflex-controlled task. However, recent research
suggests that postural stability is also achieved by high-level neural feedback
mechanisms that primarily utilize visual, vestibular, and proprioceptive sensory
information [143].
The traditional approach to building a walking robot combines classical control
systems engineering with a sophisticated hardware system which is characterized
by high performance and accuracy. It does not try to copy biological systems,
rather, it benefits from the advantages of well-proven technology and design
methods. The fundamental difference between technical and biological systems
brings up the question whether the hardware and control design of a versatile,
multi-purpose humanoid robot can actually benefit from the biological sciences.
From the author’s point of view, biological designs currently serve as a source
of inspiration. The aim is to collect suggestions for technical designs, test them
for transferability into technical systems, and realize them using state-of-the-art
mechatronics technology.
Research on humanoid robots is more and more focusing on complex, higher-
level scenarios such as human-robot interaction and manipulation. Regarding
the locomotion capabilities of biped robots, however, there are still numerous
unsolved problems. For example, higher speeds of locomotion and flexible motion
generation still remain challenging. Since of them do not necessarily require the
development of new hardware, the highest degree of attention and development
is often given to the control system, while almost no importance is attached to the
mechatronic hardware and its significance for control performance.
The review of current humanoid robot projects in Section 1.1 shows certain
“convergence” in the hardware design and component selection, suggesting that
the “ideal” architecture is found. However, research on reliable biped robots
with high performance and human-like walking capabilities still faces numerous
problems regarding the kinematic structure, mass distribution, actuators and drive
mechanisms, and sensor layout. Moreover, robots that effectively combine the
advantages of full actuation (Section 1.1.1) and the biologically inspired approach
of passively compliant structures (Section 1.1.3) is not in sight.
In conclusion, balance control and robot hardware must be seen as tightly
coupled parts of a highly integrated mechatronic system. The judicious and con-
current design of both parts is paramount to the realization of a high-performance
walking robot.
1.1 Literature Review

Two fundamentally different approaches are dominating in the bipedal locomotion
community. Besides fully actuated humanoid robots using classical control theory
1.1 Literature Review 3
© Atsuo Takanishi Lab., Waseda University
© Toyota Motor Corporation

© Honda Motor Co., Ltd.
(a) Wabian-2R (b) Asimo (c) Toyota running robot
Figure 1.1: Humanoid robots, part 1
and mechatronic design, another concept based on the idea of passive dynamic
walking is emerging. Although humanoid robots are a comparatively new self-
contained research field, the literature is tremendous. Without attempting to
review all research projects, this section describes the ones that are considered to
be the most relevant. More specific details on the servomechanisms, sensors and
electronics are given throughout this work. Aiming for a comprehensive review,
the Springer Handbook of Robotics provides two contributions on legged [85] and
humanoid [97] robots. Gienger [46] provides interesting statistical evaluations
of relevant characteristics of current robot platforms.
1.1.1 Fully Actuated Biped Walking Robots

Speaking of fully actuated systems, significant advances in enabling technologies
(sensors, drive and computer technology) has allowed the realization of numerous
humanoid robots with advanced capabilities.
The pioneering work in the field of fully actuated biped robots was achieved
by Kato at Waseda University, Japan. In the year 1973 he created WABOT-1,
the world’s first full-scale anthropomorphic robot [106]. From 1986 to 1990, a
hydraulic biped robot was developed within the WL-12 family. The robot has a
trunk and a waist with two degrees of freedom (DoFs) for more natural motions
[106]. In 1999, Yamaguchi et al. [204] presented the 15-DoF robot WL-12RVI
that employs a complex foot mechanism with elastic pads to absorb impact and
contact forces. Using these feet, the robot was able to walk over uneven terrain.
Waseda’s current robot generation, Wabian-2 (41 DoFs, size 1.53 m, mass 64.5 kg),
shows some specialties in its kinematic configuration: unlike other robots the legs
have seven DoFs with internal/external rotation axes of the ankle joints, allowing
rotation of the legs relative to the feet along the vertical axis [130]. The developers
expect improved gait pattern generation as the spatial orientation of the knee joint
4 1 Introduction
and foot is independent. Especially passing over obstacles, through narrow spaces,
or climbing a ladder are expected to be simplified. Also, Wabian-2 is equipped
with a two-DoF pelvis joint and an additional two-DoF joint in the upper body
[135]. Using the pelvis joint, the upper body can be bent forward, backward and
sidewards. The refined version Wabian-2R, Figure 1.1a, is equipped with passive
toe joints that enable walking with stretched knees and heel-off gait patterns [132].
Furthermore, Wabian-2R is able to mimic various human motions and can be used
for testing rehabilitation devices. The robot can provide quantitative information
about the effectiveness of these instruments. The top speed of the Wabian robots is
currently 0.75 km/h. Sugahara et al. [176] develop the multi-purpose locomotion
system WL-16 where the legs are designed as Stewart platforms. The robot is able
to carry a heavy payload and suitable as a locomotion platform for humans.
One of the most advanced humanoid robots is being developed by Honda
Motor Corporation. To the surprise of all research groups Honda presented the
outstanding robot P2 after 10 years of secret research and several prototypes
[65] in December 1996. P2 is 1.82 m tall and weighs 210 kg. Having a total of
30 DoFs, it was the first self-contained robot that is able to walk, climb stairs
and perform simple manipulation tasks. Besides encoders for joint position
measurement, the sensor layout comprises fiber-optic gyroscopes, inclination
sensors, ground reaction force sensors and four cameras. Its successor P3 [64]
was presented in 1997 (30 DoFs, size 1.60 m, mass 130 kg). In the year 2000,
Honda presented the robot Asimo, Figure 1.1b, which was reduced in size and
mass (26 DoFs, height 1.20 m, mass 52 kg). The control system architecture called
“i-Walk” allows the robot to change its direction of movement arbitrarily and
provides greater stability against unexpected disturbance. Besides a stereo vision
system for dynamic environment detection, Asimo integrates speech recognition
and is able to detect the speaker’s location, enabling the robot to take over simple
tasks in a human environment. The current research model Asimo-R (34 DoFs,
height 1.30 m, mass 54 kg) employs additional DoFs in its arms and hands to
extend the robot’s manipulation capabilities. At the end of 2005 Honda announced
that Asimo-R can run straight on and in a slalom course at maximum speeds of
6 km/h and 5 km/h, respectively. According to publicly available data1 , the
maximum walking speed is 2.7 km/h. As its predecessors, Asimo-R relies on
a highly accurate and responsive hardware which is improved by employing
high-dynamic servo actuators with high performance and a lightweight yet rigid
leg structure. Furthermore, the balance controller is extended by bending and
twisting of the torso. In late 2009, Honda published some details on their research
on walking and running, e. g., [185]. Using a robot with similar size as Asimo, the
authors achieve running at 10 km/h. Although many details on the mechatronic
system are not yet disclosed, these impressive results can be regarded as a proof of
concept that fully actuated robots with rigid structure are capable of fast bipedal
locomotion.
Until Honda’s new top mark, Toyota’s running robot, Figure 1.1c, used to be
the fastest humanoid robot at a top speed of 7 km/h, but walking and running is
only feasible on flat surfaces [178]. It has a total of 15 DoFs: each leg has seven
1 http://world.honda.com/ASIMO/technology/spec.html, accessed on 05/30/2010
© JSK Robotics lab., The University of Tokyo

© KAWADA Industries, Inc. / AIST
© KAWADA Industries, Inc. / AIST
(a) HRP-2 (b) HRP-3 (c) H7
DoFs including an actuated toe joint and a one-DoF pelvis joint. The robot stands
1.30 m tall and weighs 50 kg. Toyota’s research focuses on developing new forms
of mobility for elderly care and assistance robots in manufacturing and household
environments. The goal is to achieve close interaction with humans by exploiting
morphological similarities of humans and humanoids. Notably, Toyota’s running
robot is, to the author’s knowledge, the only robot that implements compliance
control for absorbing the landing impacts without force or torque measurement
in the feet and joints. Instead, controlled compliance of some joints is utilized.
The landing impacts are attenuated by adjusting the feedback gains under the
assumption of a flat ground [178].
The “Humanoid Robotics Project” (HRP) funded by the Japanese Ministry of
Economy, Trade and Industry (METI) is definitely one of the largest humanoid
robotics projects worldwide. The HRP robot hardware is developed in cooperation
of the National Institute of Advanced Industrial Science and Technology (AIST)
and Kawada Industries. The self-contained robot HRP-2, Figure 1.2a, has a total
of 30 DoFs, including six-DoF arms, one-DoF hands, a two-DoF waist, six-DoF
legs, and a two-DoF head with stereo cameras. The robot is 1.54 m tall and weighs
58 kg [95]. The angular workspace of the joints allows the robot to lay down
on the floor and get up again. The three-DoF hip joint is a specialty within the
kinematic structure of the robot: the adduction/abduction joint2 are arranged
very close to the sagittal plane of the robot, reducing the hip joint loading due
to the small distance to the center of mass. However, the contribution of the
adduction/abduction joints to increasing the step length is limited. Moreover, the
resultant hip joint stiffness is determined by the cantilevered design with a single-
side bearing of the flexion/extension axis. The top speed of HRP-2 is 2.5 km/h
[95]. Kajita et al. [87] present an experimental study of running motion using
2 see Figure A.1 on page 176 for an explanation of the anatomical terminology
6 1 Introduction
the 12-DoF leg module HRP-2P. The maximum running speed is considerably
lower than the top speed of HRP-2. In another study on running, Kajita et al. [90]
develop spring-loaded, passive toe joints for the 12-DoF leg module HRP-2LT. The
successor model prototype HRP-3P stands 1.60 m tall and weighs 65 kg including
batteries. The most important distinctions to HRP-2 are additional DoFs in the
arms and hands, making a total of 36 DoFs, and the sealed body shell, protecting
the robot against ingress of liquid, solid foreign objects and dust [3]. Furthermore,
the on-board PC unit is complemented by several decentralized computer nodes
for motor control and sensor data processing. The computers are connected by
a proprietary real-time Ethernet protocol. HRP-2’s actual successor is shown in
Figure 1.2b. HRP-3 is 1.61 m tall, weighs 68 kg and has a total of 42 DoFs [96]:
the legs have six DoFs each, head and pelvis two DoFs each, and the arms and
hands have seven DoFs and six DoFs, respectively, improving the manipulation
capabilities. As its prototype, HRP-3 has a dust-proof outer shell. The electronics
architecture is modified: now, HRP-3 has a centralized control system for the arms
and head, while decentralized nodes are applied to the legs, pelvis and hands,
using a CAN-based communication architecture [96]. In December 2009, Kaneko
et al. [93] presented the “cybernetic human” HRP-4C. Standing 158 cm tall and
weighing 43 kg (including batteries), the robot has 43 DoFs. The underlying
technology (actuation, electronics, etc.) is basically identical to previous HRP
models, but the proportions and appearance are very human-like.
The JSK lab at the University of Tokyo is another major robotics laboratory. The
former robots H5, H6 and H7 [126] were built by Kawada Industries. The robots
were planned as experimental research platforms for autonomous locomotion and
whole-body motion [41]. The joint motion ranges are adjusted such that the robots
can kneel down and climb stairs to maximum step height of 25 cm. Using the
integrated stereo camera vision system of H7 shown in Figure 1.2c, advanced path
planning methods for locomotion in unknown environments with various fixed
and movable obstacles were developed [127]. Nishiwaki et al. [128] implement
actuated toe joints in the robots H6 und H7. Walking speed increases significantly
which is mainly due to reduced angular velocities in the knee joints. One major
focus of current activities at JSK is developing technologies for multi-purpose
household assistance robots. Using several HRP-2 robots, which were partly
developed at JSK and modified in some details, methods for whole-body motion
planning and vision-based environment recognition are evaluated experimentally.
Another focus is the development of whole-body musculoskeletal humanoids
with flexible spines. The aim of this novel and—at this time—unique approach
to designing humanoid robots is to realize human-like natural motion using a
flexible spine. The “anatomy” of these robots is based on the human skeleton
and manufactured from plastic by rapid prototyping techniques. All joints are
driven by redundant tendon-shaped actuators, each of which is composed of a
tension sensor and a spring-loaded motor-gear unit. Using the tension sensors,
joint stiffness is adjustable. The latest robot Kojiro (height 1.40 m, mass 45 kg)
has a total of 82 DoFs that are actuated by 109 spring-loaded motors [124]. Kojiro,
Figure 1.3a, employs a new spherical shoulder joint mechanism [175]. The highly
redundant actuation and sensor system of these multi-DoF robots with a highly
© Institute of Applied Mechanics, TU München
© JST-ATR Computational Neuroscience Lab.

© JSK Robotics lab., The University of Tokyo
(a) Kojiro (b) Johnnie (c) CB-i
complex structure is indeed challenging, both in building and control.

The biped robot Johnnie, Figure 1.3b, was developed by Gienger [46] and Löf-
fler [109] at the Institute of Applied Mechanics, Technische Universität München,
as part of the priority program “Autonomous Walking.”3 The goal was the real-
ization of fast, dynamically stable walking on even and uneven ground. At a step
length of 0.55 m, a maximum walking speed of 2.4 km/h was achieved. Johnnie is
also able to take turns and climb up stairs. The second focus was autonomous,
vision-guided walking. Using computer vision techniques that were developed
by another institute [169], Johnnie was able to perceive its environment and to
reactively adapt its step sequences to navigate obstacles and climb stairs. Johnnie’s
proportions correspond to a male human with a body height of 1.80 m. The total
mass is about 50 kg. The robot has 17 actuated joints: each leg has six DoFs, three
in the hip, one in the knee and two in the ankle. The pelvis segment has one DoF
about its vertical axis. For compensation of the overall momentum each shoulder
incorporates two DoFs. However, the pelvis and shoulder adduction/abduction
joints were mechanically blocked and not used in the gait patterns, so that 14 DoFs
are actually used. Like most other biped robots, the joints are actuated by DC
brush motors through Harmonic Drive gears. Only the ankle joints are actuated
by parallel mechanisms with ballscrews [47]. Johnnie is autonomous to a far extent,
only energy is supplied by cables. The joint angles and angular velocities are
measured by incremental encoders on the motor shafts. Johnnie is equipped with
two customized six-axis force/torque sensors in its feet. Upper body orientation
and angular velocity are determined by a customized inertial measurement sys-
tem composed of three gyroscopes and three accelerometers. Using a centralized
on-board computer system, Johnnie is computationally autonomous. An exter-
nal PC is necessary only for monitoring purposes and to submit basic operating
3 The priority program SPP 1039 was funded by the German Research Foundation (DFG) from
1998–2003. A comprehensive overview of all projects is given in [145].
8 1 Introduction
commands in case the vision system is not connected.

The Korea Advanced Institute of Science and Technology (KAIST) is currently
working on its fourth generation of humanoid robots. The robot HUBO (develop-
ing name KHR-3) is an improved version of its predecessors KHR-1 and KHR-2
[100, 140]. HUBO weighs 56 kg and is 1.25 m tall. It has a total of 41 DoFs, includ-
ing legs and arms with 6 DoFs each and five-DoF hands. Similar to Johnnie, all core
technologies (servo actuators, force/torque sensors, inertial measurement system,
power amplifiers) are custom developments. The robot employs a distributed
real-time control system using CAN communication. The maximum walking
speed of HUBO is 1.25 km/h [139]. KAIST’s current robot HUBO2 (40 DoFs, height
1.30 m, mass 45 kg) is being developed since 2008. Each leg and arm has six DoFs
and seven DoFs, respectively. The pelvis joint has one DoF, the hands have ten
DoFs each and the head has three DoFs in the neck. According to the project
website4 , HUBO2 can walk with stretched legs at 1.4 km/h and run at 3.3 km/h.
Measuring only 0.58 m, the robot SDR-3X [83] is not full-sized, still Sony’s
research activities on humanoid robots are notable. The robot weighs only 5 kg
and has 24 actuated DoFs. As with the robot dogs Aibo Sony’s marketing activities
were focused on the entertainment area, which explains the size of the robot. The
improved version SDR-4X with a total of 38 DoFs has been presented in 2004
and later renamed to QRIO [84]. QRIO is built with modular servo actuators,
comprising a custom-developed brushless motor, planetary gears and control
electronics. These so-called “intelligent servo actuators” [40, 82] are employed
in three different sizes and seem to be optimized for mass production. QRIO can
move arbitrarily but also sit down and get up and—as is well known—it was the
world’s first running biped at a speed of 0.83 km/h [84]. In 2006 Sony announced
the discontinuation of Aibo and QRIO at the same time.
In Germany the development of the robot BARt-UH at the University of Han-
nover should be emphasized. BARt-UH (six DoFs, size 0.70 m, mass 4.5 kg) is a
planar biped without upper body. It is capable of statically and dynamically stable
walking. Each leg incorporates actuated joints in the hip, knee and ankle that are
driven by DC brush motors through Harmonic Drive gears [5]. The maximum
walking speed of BARt-UH is unknown. With a total of 12 actuated DoFs the robot
LISA (height 1.24 m, mass 40 Kg) is a lower limb robot without any actuated DoFs
in its upper body. Having three-DoF hip joints, one-DoF knee joints, two-DoF
ankle joints and passive toe joints, the robot is able to walk in arbitrary directions.
The joints are driven by DC brush motors through Harmonic Drive gears. A
specialty of LISA’s kinematic structure is the hip joint, which is designed as a
ball-and-socket joint with parallel actuation. All hip motors are attached to the
upper body [70]. The sensor system comprises six-axis force/torque sensors in
the feet and three inertial measurement systems on the trunk and thighs. LISA
employs a distributed on-board computer system using CAN communication.
The current robot BART III was presented in late 2009 [153]. The robot stands
1.25 m tall and weighs 35 kg. 19 active joints divide up into six DoFs per leg,
three per arm and one in the pelvis. BART III employs absolute angular sensors in
its joints, four load cells for measuring the ground reactions in each foot and an
4 http://hubolab.kaist.ac.kr/hubo(khr-4).php, accessed on 05/30/2010
inertial measurement system in the upper body. The robot is again controlled by a
distributed system comprising DSP-controlled power electronics and an external
main controller which are connected by CAN bus.
Although discontinued at the end of 2002, the anthropomorphic biped robot
BIP2000 (size 1.80 m, mass 105 kg) developed at the French institute INRIA Greno-
ble is worth mentioning as it features some interesting details in its mechanical
structure. BIP2000 consists of two legs with six DoFs each and a three-DoF pelvis
joint, making a total of 15 active joints [35]. All joints are actuated by DC brush-
less motors. Five joints are driven through Harmonic Drive gears, the other ones
employ linear actuators based on planetary roller screws: the knee joint and hip
joint flexion/extension axis are actuated by slider-crank mechanisms. Two-DoF
parallel slider-crank mechanisms are applied to the ankle joint and parts of the
pelvis joint. The joint angles are measured by resolvers on the motor shafts. Three
load cells per foot allow the measurement of the vertical ground reaction force
and two lateral moments. Two-axis inclinometers are mounted on the feet to
measure their inertial orientation. The orientation of all other links including the
upper body is then calculated by inverse kinematics. BIP2000 is controlled by an
on-board computer system based on VME-bus components.
A very sophisticated anthropomorphic robot with hydraulic actuation is de-
signed by Sarcos Research Corporation for ATR Computational Neuroscience
Laboratories, Japan. The robot CB is 1.58 m in height and 92 kg in mass [23]. It has
an active head with seven DoFs, seven-DoF arms and legs, a three-DoF upper body
and six-DoF hands, making 50 DoFs in total. The sensor layout is rather complex
and includes position and velocity sensing at all joints, and torque sensors on
the main joints of the arms, legs, torso and neck. Foot force sensors measure the
ground reactions and two inertial measurement systems on the head and pelvis
provide the orientations of the head and overall system, respectively. Walking
control and force-based balance control run on an on-board computer system,
any processing requiring substantial computing power (vision processing, etc.)
are performed remotely on a cluster of PCs. The robot is supplied by an external
hydraulic pump. The succeeding model CB-i, Figure 1.3c, has a total of 51 DoFs,
measuring 1.55 m and weighing 85 kg. The joints are actuated either electrically
(cameras), pneumatically (hands), or hydraulically (all other joints).
The robot PETMAN built by Boston Dynamics is another hydraulically driven
biped robot that is aimed for testing chemical protection clothing used by the
US Army. The legs look similar to Boston Dynamics’ four-legged robot BigDog.
Despite its non-anthropomorphic leg structure with four links, PETMAN has a
human-like shape and size. According to a video on the company website5 , it is
capable of heel-to-toe walking and currently achieves a walking speed of 7.08 km/h.
Pratt and Krupp [147] present the 12-DoF 3-D walking robot M2V2. The robot
is developed by Yobotics, a spinoff from M.I.T.’s former Leg Laboratory. M2V2
has the traditional leg structure with six DoFs per leg and no actuated DoFs in
the upper body. The leg joints are powered by “Series Elastic Actuators.” These
electric linear actuators are based on ballscrews and incorporate coil springs in
series between the screw output and the load. The output force is controlled
5 http://www.bostondynamics.com/robot_petman.html, accessed on 04/01/2010
10 1 Introduction
by controlling the compression of the springs which is measured by quadrature

encoders, hence the springs basically act as load cells. The authors expect to use
mechanisms that are used in passive dynamic walkers (cf. Section 1.1.3).
Some very sophisticated and unique designs of humanoid robots and its compo-
nents are developed by the Nakamura lab at the University of Tokyo. Nakamura
et al. [123] develop the robot UT-θ (height 1.50 m, mass 45 kg). UT-θ has a total
of 23 DoFs (neck three, arm four, leg six DoFs). The joints are driven by DC brush
motors through Harmonic Drive gears. The lightweight design is mainly due to the
main structure which is cast from magnesium alloy. The sensor layout comprises
incremental encoders for joint position sensing, four six-axis force/torque sensors
at the elbows and ankles and an inertial measurement system at the trunk. The
kinematic structure of UT-θ has some distinctive features: the mobility of the hu-
man shoulder is mimicked by a parallel mechanism with two-DoF prismatic and
a one-DoF rotational joint. The so-called “cybernetic shoulder” [133] allows the
continuous movement of the shoulder’s center of rotation. Considerable passive
compliance is achieved by a flexible link that can absorb impacts at a collision
with obstacles and reduce the risk of damage to objects or people. The “double
spherical hip joint” [134] is composed two spherical hip joints with shared center
of rotation. The authors expect advantages in the stabilization of the robot by
separating the upper and lower body motion such that the upper body motion
is controlled independently from foot placement. The knee joint is equipped
with a “backlash clutch” [134] that adds large backlash between the thigh and
shank. Torque is transmitted unilaterally to the shank, allowing to accelerate
the shank, while shank advancement occurs virtually passive. The current robot
UT-θ2 (20 DoFs, height 1.50 m, mass 45 kg) features basically the same design as
its predecessor, except that structural stiffness is increased and the servo actuators
are more powerful. Also, the concept of the double spherical joint [134] is adopted
to the shoulder joints, enabling a large workspace of the arms.
Other developments that should be mentioned include the HOAP robots series
[117] by Fujitsu (now Miyachi Systems), ETL-Humanoid [121], the robots BHR-1
and BHR-2 from the Beijing Institute of Technology [105], Mahru-III developed by
Samsung [104], REEM-B from PAL-Robotics [186], and the 0.57 m tall robot NAO
from Aldebaran Robotics [53]. NAO has been selected to replace the Aibo robots
in the RoboCup standard league.
1.1.2 Balance Control of Fully Actuated Biped Robots

This work does not deal with biped walking control. Since the control strategy
also affects the mechatronic design concept, this section briefly describes the most
relevant control methods. For a more comprehensive review of biped walking
control and trajectory generation the reader is referred to an article by Kajita and
Espiau [85].
In the early 1970s Vucobratović formulated the zero moment point (ZMP)
theory as a stability criterion for biped robots [196]. Since then, ZMP theory
has established as the basis for gait pattern generation and control for many
biped robots. The ZMP, r ZMP = (rx , ry , 0)T , is defined as the point in the walking
surface where the moment M due to the ground reaction force, F G = (fx , fy , fz )T ,
corresponds to the ground reaction moment, M G = (mx , my , mz )T , and is thus
identical to the center of pressure: M = M G − r ZMP × F G = (0, 0, mz )T . Note that
the condition of zero moment about the vertical axis of the contact plane, mz = 0,
is not necessary in general: this moment is compensated by the friction force as
the feet can be assumed not to slide during ground contact. The contact between
the ground and the support leg is stable if the ZMP lies within the foot support
polygon. During normal gait, it is undesirable to have the ZMP on or close to
the edges of the support polygon as an additional moment would easily cause
instability. A common interpretation of this theory is therefore to treat the distance
between ZMP and the edges of the support polygon as a measure for the reserve
of stability.
For now, a large number of approaches for balance control of biped robots have
been proposed and a number of similar ideas have been successfully implemented
on full-size biped robots. A very common approach to maintaining balance is
controlling the contact torques at the feet [89, 110]. These are usually measured
by a multi-component force/torque sensor or several load cells and controlled
via position control of the ankle joints. Another method is to accelerate the CoM
[127, 178] to create a reaction force that stabilizes the robot. Other possible
strategies utilize upper body inclination or foot positions. Reducing the landing
impacts involves the combination of passive and active mechanisms. Many robots
incorporate active control of the vertical contact forces [89, 110]. Large impacts
are avoided by using low gain joint position control in combination with gain
scheduling [127] or inverse dynamics feed forward control [178].
These control components are normally developed independently, so that un-
wanted interference is possible if they are combined. But interference can be
avoided by using an integrated control approach. Löffler [109] proposes an
integrated, model-based approach for the biped Johnnie that is based on feedback
linearization. Unfortunately, the performance of this method is limited by the
available sensor bandwidth, computational power, and model accuracy [110]. A
more fundamental issue is the fact that the computed torque method requires
joint torque control, which is difficult without joint torque sensors.
Buschmann [17, 18, 114] proposes a real-time walking control system for the
robot Lola that plans the walking trajectories on-line using a simplified robot
model. The trajectories are then modified by hybrid position/force control in task
space based on a resolved motion rate scheme. Inertial stabilization is achieved
by modifying the contact force trajectories. Comparable to most other bipeds, a
multi-layered, hierarchical control concept is implemented, with a high gain joint
position control loop on the lowest level [18].
1.1.3 Passive Dynamic Walkers

The idea of fully passive walking mechanisms that rely on precisely tuned natural
dynamics was pioneered by McGeer [116]. Descending a shallow slope without
any control and actuation but gravity, these mechanisms generate a periodic gait
with a high energetic efficiency. A more recent development are passive walking
12 1 Introduction
mechanisms which are partly actuated in few joints, thus they do not have to
walk down an incline to accommodate the energetic losses. These minimally
controlled mechanisms contain passive elements like springs and dampers that
utilize inertia and collisions for legged locomotion at minimal power consumption.
Since the natural trajectories of a passive walker are basically determined by its
mechanical structure, they are less versatile and robust than fully actuated biped
robots. Moreover, most of these walkers are designed as planar mechanisms, and
some of them are physically constrained to the sagittal plane. In consequence,
stability must be considered only in the sagittal plane since falling sideward is
not possible. Due to fundamental differences between the stabilization of planar
(2-D) and free (3-D) walking, the comparison between the approaches of passive
walkers and fully actuated biped robots is limited. Nevertheless, there are some
interesting current machines that are only partly passive.
The walkers Denise [201] and Flame [69] developed at Delft University are
“semi-passive dynamic walkers” with partial actuation. Denise (size 1.50 m, mass
8 kg) is pneumatically powered and has five DoFs in the hip, knees and ankles.
The ankles and knees are completely passive, only the hip joint is actuated by
pneumatic muscles. Denise can walk with a velocity of 1.44 km/h. The latest robot
Flame (seven DoFs, mass 15 kg, size 1.30 m) is a 3-D robot with series elastically
actuated hips, knees and ankles. Cable transmissions are employed to reduce the
mass of the legs by placing all electric motors in the torso. The actuated joints
employ series elastic elements that allow joint torque control. The joint angles
are measured by incremental encoders and the inclination of the upper body is
estimated by an inertial sensor. The control loop runs on an on-board PC and
implements torque and/or position control. Flame is able to place its feet sideways.
Currently, it walks at a speed of 1.62 km/h.
At Carnegie Mellon University Robotics Institute, Hurst et al. [74] develop the
four-DoF planar walker BiMASC to study the role of controllable compliance in
running and control strategies that utilize the leg springs. The robot is driven
by actuators with mechanically adjustable series compliance and designed to
resemble the “Spring Loaded Inverted Pendulum” (SLIP) model, a mathematical
model widely used in biomechanics. Despite structural similarities to the “Series
Elastic Actuators” used for example in the robot M2V2 (see Section 1.1.1), the
actuators of BiMASC are large enough to store the energy of a running gait,
whereas the former are primarily intended for sensitive and robust force control.
Moreover, the spring stiffness is mechanically adjustable and can be tuned for a
particular gait or ground surface.
A similar approach is investigated by Vanderborght et al. [191] at the Vrije
University Brussels. The robot Lucy (mass 30 kg, size 1.50 m) is actuated by
pneumatic muscles. Due to adjustable compliance of the pneumatic actuators, the
natural dynamics of the system can be changed according to the situation. Having
a total of six DoFs, Lucy can only walk in the sagittal plane. A guiding mechanism
prevents the robot from falling sideward.
Förg and Ulbrich [37] investigate the influence of elastic mechanisms on
technical, bipedal locomotion by the example of a 3-D multibody simulation
model of a 15-DoF robot that is equipped with rubber straps. The parameters
of the elastic elements are identified by numerical optimization using different

objective functions. The behavior of the resulting configuration is explored in
terms of biomechanics.
1.1.4 Other Relevant Developments

The broad experience of the Institute of Applied Mechanics in the development
of walking machines [144] including the six-legged walking robot Max, the eight-
legged pipe crawler Moritz, and the biped Johnnie, served as an important basis
for the current robot development.
Though originally designed for crawling not walking, the iCub project funded
by the European Commission is notable. The project aims at providing a research
platform in the area of cognitive systems, interaction and robotics. At a mass of
22 kg iCub is 1.04 m tall, corresponding to the size of a 3.5 year old child [188].
It has 53 DoFs which are mainly located in the upper part of the body. iCub can
crawl on all fours and sit up. The entire technology (i. e., mechanical structure,
electronics, software) is distributed open source.
Sophisticated lightweight robotic arms with seven DoFs are developed at the
Institute of Robotics and Mechatronics, German Aerospace Center (DLR) [66–
68]. The latest generation, LWR-III, is characterized by its modular design and
a distributed motion control architecture. The joints are driven by customized
RoboDrive motors whose characteristics are tailored to the needs of robotic appli-
cations. Local joint control is implemented on DSP boards and allows to control
output torque, position, or impedance independent. Each joint employs motor-
side position sensing, and torque and position sensing at the output. Cartesian
control runs out on a central computer which is connected to the joint controllers
over the fiber-optic SERCOS interface. In 2006, Ott et al. [137] presented the
two-arm system Justin which is composed of a newly developed upper body, two
torque-controlled lightweight robot arms LWR-III, and two four-fingered hands.
Justin has 43 actuated DoFs, including seven-DoF arms, 12-DoF hands, and a five-
DoF torso. The overall mass is 45 kg. The robot is mounted on an omnidirectional
platform. Justin is used to test programming and control concepts for bimanual
mobile manipulation.
A similar system is developed within the collaborative research center SFB-588
at the University of Karlsruhe. The main focus of this research is human-robot
interaction within a household environment. The latest robot ARMAR-III weighs
30 kg [4, 7] and has 43 actuated DoFs: the head has seven, the waist has three, each
arm and hand has seven and eight DoFs, respectively. Some drive mechanisms
have a rather complicated structure. For example, the shoulder joint is a series
of three joints driven by brushed DC motors through Harmonic Drive gears. In
addition, the shoulder adduction/abduction and upper arm internal/external
rotation joints employ worm gear transmissions. The motors for the elbow and
wrist joints are located in the upper body, driving the joints through wire ropes.
The computer system has a distributed architecture with several PCs and local
DSP modules for sensing and low-level control. The local controllers connect to
the higher-ranking controller via several CAN bus lines.
14 1 Introduction
Design and Realization

Electronics architecture,
(S. Lohmeier)
geometric constraints
– Dimensioning & design of mechanical structure
– Dimensioning & design of servo actuators
– Sensor system
– Design concept of electronics architecture
System parameter:
Fundamental design data:
Hardware Programming
topology, dimensions, (V. Favot)
joint loads, constraint
dynamic properties,
forces, external forces – Low-level motor control
motor and gear
and moments, . . . – Sensor data acquisition, signal processing
parameters, . . .
– Real-time communication system
Simulation and Control
(T. Buschmann)
– Dynamics simulation
– Trajectory generation Software interfaces
– Stabilizing control
– Load profiles for mechanical structure and drives
– Real-time communication system
Figure 1.4: Integration of the sub-project “design and realization” in the overall project
1.2 Contributions of this Thesis

This work involves the mechatronic conception, mechanical design and realization
of a dynamically stable walking robot with high performance and autonomous
walking capabilities. It is part of the project cluster PAK-146 „Natur und Technik
Intelligenten Laufens“6 High walking performance includes more natural walking
motion, human-like walking speed, flexible gait patterns for high mobility and
agility, and a high degree of autonomy. The maximum walking speed is targeted at
approximately 5 km/h which corresponds to the average walking speed of a human
but is currently achieved by only few robots. The necessary performance of the
mechanical system must be balanced with a minimal system mass. Hence the
design of servo actuators with a high power density is paramount. Autonomy of
the system is related to the drives, sensors and computer system. Since the robot
is mainly operated within a laboratory environment, energy is supplied from an
external source for the benefit of lower system mass. The robot should provide a
reliable platform for testing different control methods, including vision-guided
autonomous walking in arbitrary and realistic scenarios.
Figure 1.4 outlines the overall project organization: the robot is developed in
collaboration of two major sub-projects, “Simulation and Control” and “Design
and Realization”, which are worked on iteratively and in parallel. The third sub-
project, “Hardware Programming”, was initiated in a later project phase. The
sub-projects “Simulation and Control” and “Hardware Programming” are dealt
by Buschmann [17–20] and Favot, respectively.
A well thought-out and holistic design concept of the mechatronic system
is paramount for the robot’s dynamic performance and agility. The hardware
6 “Biological and Technical Aspects of Intelligent Locomotion.” Group website: http://www.
uni-koeln.de/math-nat-fak/zoologie/tierphysiologie/dfgGruppe.html, accessed on
05/20/2010
1.2 Contributions of this Thesis 15
development relies on the “classical” approach, using well-proven and state-of-the-

art mechatronics technology. Yet, inspiration from biological sciences is applied
wherever appropriate. The employed machine elements and design methodologies
are not generally new. Also, improving system dynamics by optimizing the
mass distribution is a well-known approach in the field of industrial robotics
and handling devices. However, global system dynamics of such conventional
manipulators differ significantly from biped robots: manipulators are normally
fixed to the environment and often have a with serial-link structure. Biped robots,
on the other hand, contact ground unilaterally constrained using their feet. Hence
the contact situation changes continuously and severely limits the amount of
forces and moments that can be transferred to the ground for propulsion and
balancing. The existing knowledge of designing fixed-base industrial robots can
therefore not be applied directly but must be adapted to the specifics of biped
robots. To the author’s knowledge there are no scientific publications on biped
robots that explicitly treat the global layout of the mechanical system (Section 2.4),
kinematic structures of individual joints (Section 2.6), and the use of parallel
mechanisms as design parameters for improving the dynamic properties of a fully
actuated biped. The kinematic structures of the knee and ankle joints proposed in
Section 2.6, on pages 53–56, aim at minimizing the resultant inertia of the legs
and are indeed novel to the field of humanoid robotics. Also, the foot design for
more effective impact attenuation (Section 2.3) is new. In summary, the key issues
of this work are:
1. Redundant kinematic configuration of the locomotor system
2. Consistent lightweight design, considering sufficient static and dynamic
stiffnesses of the mechanical structures and drive mechanisms
3. Development of a high-dynamic leg architecture with powerful yet lightweight
drives and optimized mass distribution in the legs
4. Development of a foot dynamics concept that ensures reliable ground contact
and effectively attenuates impact transients
5. High performance sensor layout to support the implementation of model-
based control methods
6. Assisting control performance from a technological point of view by dis-
tributed low-level joint control and sensor data processing
The mechatronic design concept of the robot is developed in Chapter 2, with
focus on the mechanical configuration, joint mechanisms, sensor layout, and
computer architecture. After developing the topology of the kinematic chain, the
foot dynamics concept is pointed out and general requirements on a high-dynamic
leg architecture are discussed.
Chapter 3 presents the mechanical design of the machine. The distinguishing
characteristics are the redundant kinematic structure with seven-DoF legs, the
consistent lightweight construction and a modular joint design with high power
density. The mass distribution of the legs is improved to achieve good dynamic
performance. After compiling general design rules, the torque and stiffness-based
dimensioning of the drive mechanisms is presented. The empirical approach to
designing structural components is compared to an approach based on topology
optimization and followed by strength assessment and elastodynamic analysis of
16 1 Introduction
critical components. The overall system is split into modular subsystems. The
design of these sub-assemblies is explained in great detail. The parameter iden-
tification through kinematic calibration and CAD-based link inertial parameter
estimation is described. Finally, the mass distribution is discussed and the dy-
namic performance of the implemented leg architecture is compared with other
approaches.
The sensor system described in Chapter 4 is able to acquire the complete sys-
tem state. The joint sensors for motor control and link position sensing are
explained. Specific requirements and weight restrictions demand the develop-
ment of customized force/torque sensors for measuring the ground reactions.
They are complemented by contact sensors to disambiguate the contact state of
the feet. The orientation and angular velocities of the upper body are estimated
by a high-precision inertial measurement system.
The computer system is crucial to the performance of the robot. Although
electronics hardware development and programming are not part of this work,
Chapter 5 briefly sketches the electronics architecture and its components. The
central control unit, local DSP nodes and the real-time communication system
are presented. The chapter concludes with an overview of hardware security
functions.
The experience gathered from the design and initial operation of the robot are
analyzed in Chapter 6. Based on the lessons learned useful recommendations for
future robot developments are made.
Chapter 7 finally summarizes the results and gives an outlook on future work.
2 Mechatronic Design Concept
Creativity requires the courage to let go of certainties.
—Erich Fromm
Humanoid robots are prime examples of mechatronic systems. They consist of a

basic (mechanical) system, actuators and power transmissions, sensors, control
electronics and control software. Before designing individual components, the
overall design concept of the robot is developed. This chapter explains the as-
pects that must be considered in the realization of a high-dynamic and powerful
robot with high walking performance (i. e., fast, agile and natural motions) and
autonomy, and is particularly related to the
1. Mechanical configuration
2. Design concept of the joint mechanisms
3. Sensor layout
4. Computer architecture
All aspects imply several fundamental design decisions throughout the develop-
ment process that affect the shape and performance of the overall system. Design
decisions are based on own considerations, a comprehensive study of related
literature and the results of the Johnnie project [46, 109]. To gain a deeper un-
derstanding of the requirements and constraints of the mechatronic system, a
thorough hardware analysis of Johnnie is conducted during the conception phase.
Finally, experience of the Institute of Applied Mechanics gathered from former
walking machine projects is beneficial [144].
System complexity is rather high due to the interdisciplinary nature of the
project and a high part diversity. Moreover, exacting demands on lightweight
design severely limit the use of purchased components but call for the custom
development of a broad spectrum of components, including servo actuators,
structural components, sensors and electronics. In order to cope with the huge
development effort within an acceptable timeframe, it is essential to adopt a
systematic procedure throughout the development process (Section 2.1).
The topology of the kinematic chain is developed in Section 2.2. The foot dy-
namics concept which is paramount for the walking performance is introduced
in Section 2.3, and Section 2.4 points out further requirements to achieve high-
dynamic performance of the mechanical system, including good acceleration
capabilities of the legs. The requirements on the joint mechanisms are discussed
in Section 2.5. Types of actuators and precision gears with high dynamics and
power density that satisfy the kinematic and kinetic requirements are selected.
The drive mechanisms of the leg joints are developed in Section 2.6. The sensor
layout is developed in Section 2.7 such that it is able to acquire the complete
system state, allowing the implementation of model-based control schemes. Sec-
tion 2.8 concludes with the conceptual design of the computer architecture. The
17
18 2 Mechatronic Design Concept
Requirements Product
n
Syst
atio
tegr
em D
Validation
in
esig
tem
n
Sys
Domain-specific design
Figure 2.1: General workflow in Mechanical
Electrical
the development of mechatronic Software
Control System
systems according to the guide-
Modeling & analysis
line VDI 2206 [192]
motivation for supplementing the central control unit with decentralized units
for motor control and data processing are stressed and the requirements on the
real-time communication system are discussed.
2.1 Systematic Approach to Managing Complexity

The development of a complex mechatronic system such as a humanoid robot
demands for a systematic and goal-oriented procedure. High system complexity
arises from the interdisciplinary nature of the problem and the high part count
and diversity (number of actuators, sensors, electronic components). Two different
aspects have to be considered in order to manage the project efficiently and within
an acceptable timeframe. Content-related development methods involve the
design methodology which is employed to structure the different stages of robot
development. Organizational measures comprise the timing and scheduling of
individual work stages using project planning tools throughout the development
and realization phases.
2.1.1 Methodology in the Development Process

The procedure described by the guideline VDI 2206 “Design methodology for mecha-
tronic systems” [192] forms the macro-cycle within the guiding framework of the
robot development process. It is connected to the flexible approach of the “Mu-
nich Procedural Model” which describes the planning and design process during
individual development phases [107].
Figure 2.1 shows the macro-cycle workflow of the guideline VDI 2206. It is a
general model for the development of mechatronic systems, describing the entire
range of required tools and activities. In the first step, a holistic and integrative
concept of the entire system is developed on basis of the objectives stated in the
research proposal. The robot’s physical and logical structure are described on
system-level and decomposed into sub-functions. Suitable active principles and/or
functional elements are assigned to each sub-function and the goal achievement is
2.1 Systematic Approach to Managing Complexity 19
Customer Needs Product Architecture

(Analyze Goal) (Assess Properties)
Testing and
Task and
Refinement
Objectives
(Ensure Goal
(Plan Goal)
Achievement)
Concepts
(Generate
Solution Ideas)
Specifications Detailed Layout

(Structure Task) (Make Decision)
Standard
Procedure
Figure 2.2: Overview of the Munich Procedural Model (adapted from [107])
again verified on system-level. The design process then divides into smaller devel-
opment lines. Actual implementation of the subsystems and single components is
carried out in a domain-specific manner. The development lines can be handled
independently and simultaneously, nevertheless, synchronization is important
throughout the development process. Finally, the development lines get merged
into the complete system, where the interaction of the various components is veri-
fied. The development progress must be validated continuously with respect to
the overall design concept. Thus, all phases of development are supported by the
analysis of system properties by means of simulation models, taking into account
either the complete system, subsystems, or single components. Like engineering
design in general, the robot design process is iterative and open-ended and usually
requires more than one macro-cycle.
For problem solving within the micro-cycles, the “Munich Procedural Model,”
Figure 2.2, is adopted as the underlying procedural model [107]. It can be traced
back to the four well-known steps to solving technical problems: task planning
and clarification, conceptual design, embodiment design and detail design. Al-
though a standard procedure is proposed, the network-like connections enable
iterations or recursions.
2.1.2 Project Planning

The experience from the Johnnie project suggests that at least four years should be
set for the design and realization of a complex humanoid robot and the necessary
control software. Organizational planning of the project is the second element
to keep track of progress and to better cope with complexity in the development
process. The first step is to develop the work breakdown structure of the project.
The common approach in product development is to use an object-related WBS.
The product structure is defined, necessary work packages are shaped, logical
and/or chronological dependencies among work packages are identified, and

the activity duration is estimated. All work packages are then transferred into
a Gantt chart of the project. The activities are represented by time-scaled bars
and precedence relationships between them are shown as arrows. Besides the
scheduling of activities, resources can be coordinated, e. g., manufacturing in the
in-house machine shop, or the acquisition of procured items.
2.2 Kinematic Structure

The first tasks in the robot design process [163] are to determine the topology of
the kinematic chain and the geometric dimensions of the links underlying the
mechanical structure. The geometric proportions of the robot structure presented
in Section 2.2.1 correspond to a human adult. The kinematic chain is derived
in Section 2.2.2 using a well-known model of human gait. Finally, the expected
advantages of the newly introduced DoFs (head, arms, pelvis, toes) are described
in Sections 2.2.3 to 2.2.6.
This thesis relies on the same reference planes used in human anatomy and
biomechanics to describe body regions and the relative movement of body struc-
tures. A short overview of the anatomic and biomechanical fundamentals of
human locomotion is given in Appendix A. Furthermore, the Rancho Los Amigos
Convention is used to divide the gait cycle into stance (loading response, mid-
stance, terminal stance, pre-swing) and swing (initial swing, mid-swing, terminal
swing) phases, see for example [142].
2.2.1 Geometric Dimensions of the Links

The geometric proportions are based on the human anatomy as far as possible.
There are numerous studies and statistics on anthropometry. Henze [63] ag-
gregates statistical data about the distribution of body dimensions from surveys
by NASA and the German Institute for Standardization (DIN) [30, 31] in the
computer program calcman3d. Specifying gender, and body size and mass, the
software determines an average link geometry and mass and inertia parameters.
The measurements are based on these sources and correspond to an adult that
stands 180 cm tall and weighs 75 kg.
2.2.2 Topology of the Kinematic Chain

Though published over 55 years ago, the “Major Determinants of Human Gait”
defined by Saunders et al. [162] still guide many researchers’ thinking about hu-
man gait. Based on the hypothesis that minimizing vertical and lateral movements
of the center of mass (CoM) during gait minimizes the metabolic cost, the authors
identified six basic motion patterns which reduce and smooth CoM vertical and
lateral movement: Pelvic rotation, pelvic obliquity (i. e., motion in the coronal
plane), knee flexion in the stance phase, knee and ankle interaction and lateral
pelvis displacement. Recent investigations, however, raise the question whether
the reduction in CoM displacement actually results from these six determinants.
2.2 Kinematic Structure 21
z
R
A
x y
F
Figure 2.3: Kinematic scheme of the robot (R=internal/external
rotation, A=adduction/abduction, F=flexion/extension)
Gard and Childress [43] find that pelvic obliquity and knee flexion do not
reduce CoM vertical displacement as much as originally described, but they are
believed to provide shock absorption during loading response. Croce et al. [25]
quantified the contribution of pelvic rotation in reducing CoM vertical displace-
ment to only 10 %, suggesting that pelvic rotation is only a minor determinant
of gait. But the authors state that heel rise during terminal stance is probably
the most important motion in reducing CoM displacement. Nevertheless, Saun-
ders’ “six determinants of gait” provide a clear and useful basis to determine
the leg topology. Moreover, the hypothesis of minimal CoM motion is in line
with established biped walking control strategies: most of them are based on the
assumption that the CoM is kept at a constant height to decouple its fore-aft and
lateral dynamics [88].
The goal of this research is the development of a humanoid robot with human-
like proportions, range of motion, and kinematic structure. The robot should be
able to move in arbitrary directions with human-like gait and speed of locomotion.
The topology is thus mainly determined by the characteristics of human walking.
Starting from the topology of Johnnie, a suitable joint structure is developed, taking
into account literature on human gait analysis and coordination. Since the number
of actuated degrees of freedom considerably influences the robot’s complexity
and weight, Johnnie achieves the desired mobility with a minimum number of
DoFs [46]. Although this approach still applies to the new design, simulations and
experiments suggest that additional DoFs in a redundant configuration allow to
implement more natural and flexible gait patterns and, more generally, extend
the abilities of the robot. Moreover, kinematic redundancies can reduce the joint
loads, provided that appropriate algorithms for redundancy resolution are used.
The kinematic scheme for the robot is therefore planned with 25 actuated DoFs,
forming a tree-structure topology with the upper body as the root link. Figure 2.3
shows the proposed robot topology. The trunk has two and the head three joints,
the arms have six joints (three each) and the legs 14 joints (seven each). Regarding
the joints of leg and pelvis, the kinematic chain of the upper body with respect to
the stance foot is redundant with nine DoFs.
During experiments with Johnnie it became evident that testing new controller
Table 2.1: Angular workspaces of the joints

Joint # DoFs Axis Workspace [°]
Head 3 Pan ±115

Tilt -15∼55
Vergence 0∼50
Shoulder 2 Flexion/extension -180∼45
Adduction/abduction -180∼5
Elbow 1 Flexion/extension -150∼5
Pelvis 2 Internal/external rotation ±30
Adduction/abduction ±15
Hip∗ 3 Internal/external rotation ±25
Adduction/abduction -50∼22.5
Flexion/extension -110∼45
Knee 1 Flexion/extension -5∼125
Ankle∗ 2 Adduction/abduction ±35
Dorsiflexion/plantarflexion -62∼45
Foot (toe) 1 Flexion/extension -65∼5
Total 25
∗ The hip and ankle joint workspaces are non-rectangular
concepts and gait patterns can benefit from larger workspaces. For instance,
moving over obstacles and climbing stairs were not considered in the design and
limited the robot’s agility to smaller obstacles and steps. Moreover, the maximum
step length was limited by the workspace of the ankle joints. For high mobility
and agility, the joint motion ranges are therefore enlarged. Table 2.1 lists the
angular workspaces of the joints.
2.2.3 Camera Head

As the robot should be able to cope with complex and realistic scenarios at consid-
erable walking speed, close interaction of the vision and path planning system and
the balance controller is necessary. The head DoFs are integrated to the controller
structure of the overall system so that the kinematic redundancies can be utilized
advantageously for gaze control.
Most humanoid robots have rather simple head designs where a stereo camera
pair is mounted on a two-DoF pan-tilt unit, such as Johnnie [26] and the HRP robots
[95, 96]. The camera head of Johnnie was operated as a self-contained unit within
the vision and path planning system. Gaze direction was controlled independently
from the robot posture. A patent application filed by Honda proposes two-DoF
pan-tilt unit, where the field of view in the tilt axis is extended by a four-link neck
mechanism [207]. Some humanoid robots show more complex designs. HUBO
has a six-DoF head, with two DoFs in the neck and independent pan-tilt axes for
both cameras [139]. The head of iCub has six DoFs, too, but the neck has three
DoFs. The cameras have independent pan axes but share a common tilt axis [11].
ARMAR-III has almost the same configuration but a four-DoF neck, making seven
DoFs in total [7]. CB has seven DoFs with three DoFs in the neck. Each camera is
actuated by a two-DoF parallel mechanism, enabling independent pan and tilt

movements [23].
Design Approach
The stereo camera pair for the new robot is mounted on a two-DoF pan-tilt
platform, enabling both smooth pursuit and saccadic camera motions. The stereo
basis is manually adjustable. In addition, camera vergence (the simultaneous
rotation of both cameras in opposite directions) is adjustable, enabling stereo
vision of objects close to the robot.
2.2.4 Arm Joints

Most humanoids have six or seven-DoF arms and hands for manipulation and
whole-body activities [97]. Since this research is focused on bipedal locomotion,
the arm structure must enable the compensation of total body momentum and is
kept as simple as possible.
Arm motion in human gait occurs spontaneously in opposition to the legs [142].
Especially at higher walking speeds arm swing provides a meaningful counterforce
to leg momentum and contributes to reducing the total body momentum about the
fore-aft axis and avoiding stance foot slippage [16]. A recent study by Umberger
[190] on the importance of arm swing for gait kinematics, kinetics and metabolic
cost, confirmed the function of arm movement. While joint angles, angular
velocities, and ground reaction forces are almost identical for walking with and
without arm swing, some kinetic variables—in particular, the vertical momentum
between foot and ground—show greater differences for walking with and without
arm swing.
Johnnie has arms with two-DoF shoulder joints, but the adduction/abduction
joints are blocked and not used in the gait patterns. The elbow joint is hinged
and manually adjustable. The arm endpoints accept different balance weights,
allowing to vary mass and inertia [46].
Design Approach
Each arm is equipped with three actuated DoFs, including flexion/extension and
adduction/abduction joints in the shoulder and flexion/extension in the elbow.
Since manipulation capabilities are not considered within this project, the arms
are planned with balance weights at their endpoints. Introducing elbow joints is
advantageous for higher walking speeds as they permit translational movement of
the arm endpoints which is also observed in human gait [142]. Translational arm
swing allows for an effective compensation of the total body momentum about
the vertical axis because it does not affect the normal forces of the foot-ground
contact. Though not relevant in straight walking, arm motion in the coronal plane
is important for arbitrary movement and makes the gait look more aesthetic and
natural. For example, the two-DoF shoulder joints can be utilized during sideward
walking to avoid collisions of the arms and legs while compensating the total body
momentum. Moreover, the resultant arm inertia is dynamically adjustable using
the elbow and shoulder joints. Using the elbow joints, the position of the arm
endpoints can serve as an additional parameter for controlling the CoM height.
2.2.5 Pelvis Joint

During normal gait the pelvis moves asynchronously in all three dimensions but
only a very small motion arc accounts for pelvic flexion/extension compared to
pelvic adduction/abduction and internal/external rotation [200]. Humans utilize
pelvic rotation to lengthen the step at walking speeds above 3 km/h, which is also
referred to as the “pelvic step” [16]. Moreover, upper body rotation is another
parameter to compensate the total body momentum. In human walking, however,
the contribution of upper body rotation is small compared to the contributions of
the legs and arms [16].
Johnnie has a one-DoF pelvis joint which is blocked and not used in the gait
patterns. HRP-2 [95] has a two-DoF pelvis joint with flexion/extension and inter-
nal/external rotation joints. This configuration is especially useful for extending
the workspace of the arms, e. g., when reaching distant objects. Also, it enables
the robot to stand up from lying on the ground [39]. Omer et al. [135] propose
a two-DoF pelvis joint combined with a two-DoF upper body joint for Wabian-2.
The kinematic redundancy makes pelvis movement independent from upper body
orientation, thus the robot is able to grasp something (e. g., a handrail), or to lean
against an object without causing a statically indeterminate situation.
Design Approach
Neglecting pelvic flexion/extension, a joint with two actuated DoFs in the adduc-
tion/abduction and internal/external rotation axis is implemented between the
trunk and the pelvis. Hence upper body orientation becomes almost independent
from pelvic orientation. The pelvis joint allows to lengthen the steps and extends
the robot’s motion capabilities for sideways walking and other motion patterns.
Moreover, the adduction/abduction joint enables walking with a stretched stance
leg, while CoM is kept at an almost constant height [142].
2.2.6 Toe Joint

A common approach to gait generation for biped robots is to prescribe the upper
body position and orientation and to calculate the joint angles accordingly. In
general, six actuated joints per leg are sufficient to define the six rigid-body DoFs
of the trunk. Hence most existing full-size biped robots have six actuated DoFs
per leg—three in the hip, one in the knee and two in the ankle joint. The feet are
usually one-piece segments without further segmentation, cf. Section 2.3. Higher
speeds of locomotion are generally achieved by lower cycle times and/or larger
step lengths. The cycle time is mainly restricted by actuator performance, and
step length limitations arise from the leg topology, joint workspaces and stance
leg singularity at initial contact. An approach to further enlarging the step length
is rolling the feet from heel to toe. However, heel lift-off with monolithic feet
is barely feasible: the robot is underactuated due to the line contact of the foot
leading edge and the floor and, thus, marginally stable.
In human walking the toes play an important role in enlarging the weight-
bearing area of the feet and are in contact for about 75 % of the gait cycle. During
the gait cycle the center of pressure moves from the heel to the toes, so that
about 40 % of the body weight is imposed on the toes in the final gait phases
[73]. As figured out by Kerrigan et al. [98], heel rise in human gait contributes
a considerable portion of the reduction in CoM vertical displacement. Heel rise
occurs during terminal stance, shortly before the swing leg touches ground [142].
An experimental study by Takao et al. [180] with humans wearing mechanically
constrained shoes shows that one-segment feet are clearly inferior to feet with toe
segments both for straight and slalom walking. Typical models of human gait
treat the feet as rigid links and provide six DoFs per leg, combining all movements
of the toes and other foot joints into the ankle joint. Since the anatomical structure
of the human foot is by far more complex, such simplifications are not valid for
faster locomotion and running [129]. A more detailed model proposed by Carson
et al. [22] separates the foot into three links: the hindfoot, forefoot and hallux, the
latter representing the toes.
Due to the importance of stance leg heel rise in human walking, the idea of
implementing toe joints on a humanoid robot is not new. One can basically dis-
tinguish between robots with passive and active toe joints. A simulative study
by Sellaouti et al. [170] proves the relevance of toe joints for step length aug-
mentation. By comparing gait patterns for flat and toed feet, the authors increase
the walking speed by 50 % for feet with passive toes. Ogura et al. [132] im-
plement passive toe joints in the robot Wabian-2R to walk with stretched knees
and heel-off gait patterns. Kajita et al. [90] develop spring-loaded, passive toe
joints for HRP-2LT to retrieve kinetic energy during running. The advantages of
passive toe joints are obvious: they are simple and lightweight in design, and heel
rise occurs automatically due to the whole-body dynamics. Like monolithic feet,
however, heel-off with passive toe joints cause an underactuated phase at the end
of single support because moments about the lateral axis cannot be transferred
to the ground. These problems are overcome by actively driven toe joints. But
the foot design gets more complex and the need for a motor-gear unit increases
the weight of the foot significantly. To the author’s knowledge, there are only
three full-size humanoid robots with active toe joints. The robots H6 and H7 use
actuated toe joints to reduce knee joint velocity during double support phase.
In an experiment with H6, Nishiwaki et al. [128] increase the walking speed
by 80 %. Toyota’s running robot [178] is able to run as fast as 7 km/h by using
active toe joints. Yamamoto et al. [206] develop a toe joint for the miniature
humanoid robot UT-µ2 using a parallel four-bar mechanism. The robot contacts
ground with the toe tips and a non-movable part of the toe segment which allows
to decrease the required joint torques and the actuator mass. The kinematics of
the mechanism yield a foot displacement in the fore-aft direction that depends on
the toe angle and the radii of the contact elements and must be taken into account
in the trajectory planning algorithms.
Design approach
For reduced joint loading and larger step lengths, an additional link is imple-
mented between the forefoot and heel, equivalent to the human toes. Figure 2.4
illustrates the benefits of actuated toe joints. The major advantage is the larger
step length. Heel lift-off is feasible and allows the swing leg to be in a more
extended configuration during mid-swing. The actively controlled area contact of
(a) Increased step length (b) Higher mobility (c) New motion
patterns
Figure 2.4: Advantages of actuated toe joints to motion generation
the toe segment allows to transfer stabilizing moments about the axes tangential
to the ground. Moreover, the joints facilitate rolling the feet from heel to toe which
reduces the joint loads in the hip and knee, as opposed to the conventional six-DoF
leg configuration. In general, actuated toe joints augment the robot’s mobility, it
is able to climb higher steps, and can execute more complex, or all new motion
patterns.
2.3 Foot Dynamics

The feet are key components in the mechanical structure of the robot. During
normal locomotion, they are the only parts of the robot which are subjected to
external loads. At initial contact they must withstand shock loads of up to a
multiple of the robot’s weight. Since the stabilizing controller is based on the
assumption that the robot does not slip and that tangential contact stiffness is high,
the feet must ensure proper ground contact during the support phases in order to
effectively transmit the necessary stabilizing forces and moments. The feet are the
end links of the legs and undergo the largest accelerations. The design concept
must therefore put emphasis on the mechanical design to be as lightweight as
possible, yet robust enough to accept high impact loading.
Despite the obvious fact that the foot mechanism has a strong influence on the
walking performance, there is surprisingly few detailed literature. Yamaguchi
et al. [205] develop a complex foot mechanism for the biped robot WL-12RVI. It
has four discrete contact points and is able to both attenuate landing impacts and
adapt to unknown, uneven floor. The elastic layers are made from an open-cell
foam material with nonlinear properties. Linear potentiometers measure the
deformation of the feet. An additional passive DoF, whose configuration depends
on the contact state of the foot, stabilizes the walking motion. Although some
current robots have divided feet with toe joints, see Section 2.2.6, monolithic
feet are the state-of-the-art in full-size humanoid robots. They normally consist
of a monolithic base plate covered with rubber or plastic to increase friction.
The feet have either four discrete contact points or area contact with the ground.
Most often compression/shear mounts are used for shock absorption [64, 94]. The
2.3 Foot Dynamics 27
stiffness of these elements under shear loading is considerably lower than under
pressure loading which limits the bandwidth of transferable forces and moments
and, thus, the performance of force control. Moreover, the material properties of
these rubber-metal elements is mainly elastic but damping is rather low. The shear
stiffness is increased in a refined version of the HRP-2 foot by a combination of
compression/shear mounts with ring mount elements [136]. Honda filed several
patent applications on design features of humanoid robot feet. The foot structure
of the robot P3 holds several elastic elements for shock attenuation. Position
sensors measure the displacement of the outer sole with respect to the main
foot segment [182]. Another approach improves shock absorption by means of a
reinforced elastomer layer with viscoelastic insets at the heel [49]. A more recent
patent application claims an ultra-lightweight foot design with plate springs made
from fiber-reinforced plastic and damping material, assembled in a sandwich-
like structure [184]. Another foot system which is adaptable to uneven terrain is
developed by Hashimoto et al. [62]. The contact elements are attached to each
corner of the foot and act as vertical sliding mechanisms. The sliders are locked
after they are adjusted to a surface. Gienger et al. [47] develop a foot structure
for the robot Johnnie which is similar to the foot mechanism of WL-12: it has four
chamfered ground contact elements coated with rubber. A large block-shaped
damper made from viscoelastic material is integrated between the upper and
lower foot plates to reduce landing impacts. Linear potentiometers measure the
deformations of the damping layer. A passive rotational DoF between the leading
and rear contact elements makes the foot-ground contact statically determinate
and allows to compensate smaller unevenness. However, the damper block is
rather compliant under shear and pressure loading which easily causes instability.
In particular, deformation of the damper by lateral moments considerably limits
controller bandwidth. Therefore, this design was replaced by a simplified foot
system with four point-contact elements. Each contact element comprises a
viscoelastic damping layer and a high-wear contact layer. Shear of the contact
elements due to tangential forces is omitted by a sliding contact bearing, allowing
only vertical compliance.
2.3.1 Biomechanics of the Human Feet

Due to numerous conflicting requirements, the development of a functional foot
design is not trivial. Hence the main functions of the human foot are discussed
before specifying the requirements on the robotic foot. Without oversimplifying
the foot motion during walking, the complex anatomical structure of the human
foot can be reduced to three functional units: (a) hindfoot, (b) forefoot, (c) hallux
[22]. Key elements are the toe joint discussed in Section 2.2.6, the subtalar and
midtarsal joints [203], and the heel pad [157].
It is obvious that several mechanisms—both active and passive—are involved to
protect the human body from the large impact forces during initial contact [199].
Active mechanisms are muscle contraction and joint movement1 at initial contact.
It is supposed that the use of knee flexion and ankle joint plantarflexion is not to
1 In particular, knee flexion, ankle joint plantarflexion, and pronation of the subtalar joint.
reduce the impact force itself but the rate at which body weight is transferred to
the stance foot [199]. The motion of the subtalar joint lowers the ankle and leg,
increasing the time over which the foot is decelerated and energy is dissipated
[13]. Due to their reaction time, however, active mechanisms can only attenuate
impacts of a certain bandwidth. Passive shock attenuators are body tissues in the
heel pad and the joints with viscoelastic properties, acting as low-pass filters by
absorbing higher-frequency components of the transient forces [146]. The heel pad
is between 10 mm and 20 mm thick and spreads the load over the whole plantar
surface [157]. The compression during initial contact ranges from 3 mm to 8 mm
[199].
During initial contact and early stance, the human foot is required to be a flexible
and adaptable structure to accommodate to unevenness in the terrain. Conversely,
during the propulsive phase it must perform as a rigid lever to transmit effectively
the propulsive forces which are applied to it. These apparently contradictory
functions are accomplished by a complex interaction of the subtalar and midtarsal
joint [13]. In addition, the progressive material properties of the heel pad tissue
ensure ground contact during stance [157].
2.3.2 Design Concept of the Robotic Feet

To ensure the effectiveness and performance of force control, the feet must be able
to reliably transmit the contact forces and moments and provide high grip on
different surfaces. The impact forces at initial contact can range up to a multiple
of the body weight. In addition, higher impact forces result from disturbances
like structural vibrations of the leg structure, or kinematic or positioning errors.
Such errors let the ground contact of the swing leg occur too early or too late
and can cause instability of the robot. Therefore, certain compliance of the foot
structure must provide essential shock absorption to enable a smooth and stable
walking motion. Adaptation to uneven terrain, on the other hand, may require
complex mechanisms such as the human subtalar/midtarsal joint complex [13].
However, the ability to adapt to rough and inclined surfaces is not mandatory
in a first version of the foot design because most experiments are conducted
within the laboratory environment. Instead, it is sufficient to compensate smaller
irregularities of the ground.
Furthermore, there are some requirements on top of the properties of the human
feet. Empirical testing and modification are important steps in the development
of a sophisticated foot design. Moreover, the mechanical structure must be mod-
ifiable and expandable, enabling the experimental validation of different foot
concepts. For an easy replacement of the heel and toe segments the feet must show
a modular structure. Finally, the systematic lightweight design is crucial as the
feet are the end links of the leg kinematic chain. In summary, the requirements on
the robotic feet are:
– Provide high grip on different surfaces to ensure proper ground contact
– Rigid mechanical structure to transmit the propulsive forces at a bandwidth
sufficient for system stabilization
– Compensate smaller unevenness
2.3 Foot Dynamics 29
1
Heel position after
initial contact 7
13
12 6 5
4 9
9 8 8
10 10
11 11
2 3
1 Forefoot 6 Force/torque sensor 10 Viscoelastic layer

2 Heel 7 Ankle joint flange 11 Wear-resistant outsole
3 Hallux 8 Contact sensor 12 Hydraulic shock absorber
4 Heel joint (passive) 9 Contact element (rigid) 13 Mechanical end stop
5 Toe joint (active)
Figure 2.5: Design concept of the robotic feet. The passive heel segment and hydraulic
damper account for effective for shock attenuation at initial contact.
– Passively compliant elements with good damping characteristics for shock

attenuation at initial contact
– Progressive spring characteristics at ground contact ensures effectiveness of
stabilizing control
– Minimal weight
– Keeping the anthropometric dimensions, if possible
– Modular mechanical configuration, simplifying the replacement of heel and
toe segments and/or other components
Based on these requirements, the design concept of the robotic feet is derived as
shown in Figure 2.5. The foot is composed of three segments with the forefoot (1)
as the base link. The heel (2) and hallux (3) segments are connected to the forefoot
(1) by the heel joint (4) and the toe joint (5), respectively. The heel joint (4) is a
passive joint that corresponds to the subtalar joint of the human foot [13]. The toe
joint (5) is actuated in the original concept, but the modular mechanical structure
facilitates different foot concepts, for example monolithic feet or feet with passive
toe joints. The force/torque sensor (6), cf. Section 2.7.3, is an integral part of the
forefoot (1) and interfaces to the ankle joint flange (7). Each contact element of the
heel and hallux holds a contact sensor (8) to acquire the contact state of the feet,
cf. Section 2.7.4. The contact forces and moments are transferred to the ground
by area contact of the contact elements (9–11) for improved stiffness of the foot
structure, compared to designs with point contact elements. Each contact element
is composed of the rigid base body (9), the viscoelastic layer (10) and the outsole
(11). The viscoelastic layer (9) corresponds to the fat pads in the human heel and
toes [146]. It shows good damping characteristics, no significant compression set,
long lasting ability to provide shock attenuation, and certain elasticity to recover
from deformations. The outsole (11) is made from wear-resistant, high-friction
rubber. Both layers are connected without any rigid part in between. Thus, the
feet interact with the environment only via viscoelastic elements and can easily
adapt to the ground surface by deformation.
Several mechanisms are combined for effective shock attenuation at initial con-
tact, acting as mechanical low-pass filters for the transferred forces. Besides the
viscoelastic contact elements (9–11), the passive heel segment (2) and hydraulic
shock absorber (12) account for additional energy dissipation, similar to the hu-
man hindfoot. The mechanical end stop (13) determines the heel position after
initial contact and thus ensures proper transmission of the propulsive forces and
stabilizing moments during ground contact. Although necessary for the perfor-
mance of force control, the compliant foot-ground contact bounds the mechanical
bandwidth of the ground reaction forces and moments. The mechanisms for pas-
sive shock absorption are thus limited by the bandwidth required by the balance
controller and must be complemented by active mechanisms. Possible approaches
are vertical impedance control of the foot touching the ground [18], or joint servo
gain scheduling of the swing leg [127].
2.4 General Requirements on a High-Dynamic Leg

Architecture
Without doubt, many open questions make the control of fast bipedal locomotion
very challenging. High dynamic stability and robustness against disturbances are
the key mechanisms to maintain balance in different situations. The integrated
control approach developed by Buschmann [18, 114] combines the stabilization
of global dynamics by manipulation of the contact forces and hybrid force/position
control in task space.
On the other hand, it is often neglected that the performance of balance control
also relies on the mechatronic hardware. While considerable effort is paid to
the development of powerful computer architectures (cf. Section 2.8), the effects
of mechanical system dynamics are often disregarded in the design concepts of
biped robots. Indeed, sophisticated control strategies may compensate for some
limitations of the mechanical structure, and sophisticated planning algorithms are
able to optimize the trajectories, e. g., for minimal speed and acceleration patterns,
and take into account constraints, such as saturation effects of control variables.
However, principal limitations of system performance can arise from fundamental
reasons, that is, reasons based on the physics of a poor mechanical architecture.
For example, fast motions of the swing leg produce large vertical contact moments
in the stance foot that can easily cause slippage in the foot-ground contact, if the
mass distribution in the legs is disadvantageous and—accordingly—leg inertia is
high.
“Dynamic performance” of fixed-base robotic manipulators is defined as tra-
ditional performance criteria, e. g., actuator limits [6]. Regarding the specifics of
legged robots, further performance limitations arise from the constraints asso-
ciated with unilateral foot-ground contact. Viscoelastic contact elements at the
feet make the robot underactuated and put severe limitations on the transferable
2.4 General Requirements on a High-Dynamic Leg Architecture 31
∆yCoM
r tc
∗
r tc
CoM
L̇
r sc
∗
r sc
zCoM
(a) Stiffness (b) CoM height (c) Link and leg inertia
Figure 2.6: Design objectives for improved dynamic behavior of the locomotor system
at higher speeds of locomotion. Left to right: high effective stiffness; maximal height
of the robot’s center of mass; low-inertia design of the leg links.
forces and moments necessary for propulsion and to stabilize the upper body
DoFs. In summary, the most important indices for evaluating the dynamics of
fully actuated biped robots are:
– Mobility (topology of the kinematic chain, space of motion of the legs and
arms, degree of redundancy, cf. Section 2.2)
– Agility (speed and acceleration capability)
– Force capability (magnitude of propulsive and stabilizing forces bounded by
the compliant unilateral ground contact, cf. Section 2.3)
– Actuator limits (joint torque limits)
– Mechanical properties (structural stiffnesses, natural frequencies, mass,
amount of moving masses)
– Accuracy
The acceleration capability of the locomotor system is closely related to the
actuator limits and mechanical properties. Evidently, lightweight mechanical
structures contribute to better dynamic properties, which are usually penalized
by added weight. Moreover, they allow for smaller servo actuators, which in turn
yield additional weight savings.
Focusing on stiffness and damping properties, Rivin [154] addresses critical
aspects of the mechanical design of fixed-base robotic systems. Most of these
approaches can be adapted to the specifics of biped robots. In addition, three
basic design objectives aiming at high-dynamic behavior of the locomotor system
are illustrated in Figure 2.6 and discussed in the following subsections: (a) high
effective stiffness; (b) maximal height of the robot’s total CoM; (c) low-inertia
design of the leg links.
2.4.1 High Effective Stiffness

The effects of structural link flexibility and joint elasticity are often neglected
under the assumption of a “rigid” mechanical structure. However, in terms
of the above “traditional” criteria, mechanical compliance (whether introduced

intentionally or not) normally degrades the robot’s performance. For example, link
deflections and joint compliance arising from the lightweight and slender design
of the mechanical structure can severely limit the static (steady-state error) and
dynamic (structural vibrations, deviations from desired trajectories) performance
of the locomotor system. Further, if the foot-ground contact occurs too early
due to mechanical compliance, Figure 2.6a, the walking performance degrades
and may at worst destabilize the robot. The problem gets more emphasized at
higher walking speeds, as the propulsive forces and landing impacts increase
significantly.
Only few publications deal with effects of structural compliance on the walking
performance of fully actuated bipeds, which makes it difficult to find reference
values for the required stiffnesses. Carbone et al. [21] analyze the elastostatic per-
formance of the robot Wabian-RIV. The robot is modeled as an elastic multibody
system using linear and torsional springs. Neglecting backlash of the gears and
joint bearings, the stiffness matrix is determined experimentally by measuring
the linear and angular displacements for known static wrenches applied at the
feet. According to the explanation of the experimental procedure, it seems that
the joints were position-controlled during the experiment, hence, the performance
of the joint position controllers is reflected in the stiffness parameters. Kwon et al.
[104] conduct a numerical modal analysis to estimate the elastodynamic behavior
of individual components, however, the dynamics of the overall system are not
considered.
Unlike conventional robotic manipulators the locomotor system of a biped
robot is subjected to topology changes throughout the walking cycle, depending
on the contact state with its environment: during single and double support the
legs respectively form an open and closed kinematic chain. Further, alternating
kinematic configurations of the stance and swing phases, Figures 2.7b and 2.7c,
put different requirements on the links and joints. Moreover, resultant load-side
inertia of the stance leg drives is significantly higher due to contributions of
upper body and swing leg. It is important to take these effects into account when
developing the stiffness and weight requirements on the leg joints and links.
Structural Compliance
The legs of a biped robot normally are serial-link mechanisms with cantilever
structure. Each link pivots about a transverse axis at one end, relative to the
preceding link. Using the analytical model of a cantilevered beam, Rivin [154]
discusses general requirements on links of industrial robots that rotate about a
transverse axis. In general, such links require (a) high stiffness in both bending
and torsion to reduce deflections of the feet due to gravity and inertia forces;
(b) high natural frequencies to achieve fast rate of decay of vibrations caused by
transient motions. Figure 2.7a illustrates shear Q, moment m and displacement u
of a cantilever beam under constant angular acceleration. Consequently, a well-
designed link tapers in cross section and/or wall thickness to reduce gravity and
inertial loading. However, the need to provide space for electric cables, actuators,
power transmission components, etc., as well as space requirements for the joint
connecting the succeeding link often limit the use of variable cross sections.
v1
x
ϕ̈stance
l 1c
E, I , %, A v2 v3 khip
L ϕ̈
l1
m1
Q
kknee ϕ̈stance
l2c
L3 m m2
m0 = 3 %Aϕ̈
l2
ϕ̈swing
u kcontact
kankle z
z mfoot x
y
(a) Cantilever beam (b) Swing leg (c) Stance leg
Figure 2.7: Requirements on the leg structure for high stiffness. Left: elastic cantilever
beam; right: changing kinematic configuration of the legs during the swing and
stance phases.
For an industrial manipulator arm, these results suggest high stiffness of the
proximal link, while low weight is more emphasized for distal links. In the case
of a biped robot with structure-varying topology, however, this design approach
would at a first glance alternately apply to the thigh and shank links. From robot
simulations, however, it can be seen that the highest link accelerations occur in
swing phase, hence this phase puts higher demands on leg dynamics than stance
phase, where the load-bearing capability is more relevant. Moreover, since a
timely ground contact of the stance leg is critical for walking stability, minimizing
deviations from the desired swing leg trajectory is an important design goal.
Although the roles of material, cross-sectional properties and support in altering
the fundamental frequency are limited, the influence of these parameters on lateral
vibrations can be studied using the dynamic beam equation:
2 2 u 2 u
!
EI + %A = f (x) (2.1)
x2 x2 t2
Here E is the Young’s modulus, I the second moment of area, A the cross-sectional
area, % the material density and f (x) an external load. For a uniform beam,
E, I, %, A = const., of length L the natural frequencies for free vibration, f (x) = 0,
can be computed from Eq. (2.1) as
s
EI
ωi = λ2i . (2.2)
%AL4
Here the eigenvalues λ2i relate to the boundary conditions of Eq. (2.1) and the
particular mode i of interest. The corresponding mode shapes, v i , for the first
three solutions are illustrated in Figure 2.7a.
The relationship between beam support, material and geometric parameters and
the fundamental mode of vibration can be seen from Eq. (2.2). While λ2i could only
be increased by providing additional support, the cantilever support naturally
limits beam stiffness. Length L is constant and given by anthropometric data,
so that the only parameters for maximizing the natural frequencies are material
and geometry. The material properties, denoted by specific modulus E/%, suggest
the use of high specific modulus materials, e. g., fiber-reinforced composites (cf.
Section 3.10). Beam cross-sectional geometry is represented by I/A. This ratio
is evidently maximized for geometries showing high second moment at small
cross-sectional area. In other words, the links should be designed with large,
closed cross sections and minimum wall thickness.
Joint and Drive Compliance
Joints, actuators and drive mechanisms are critical design components, since they
typically represent the largest part of total compliance of robotic structures [154].
Optimization of effective stiffness must therefore include compliance of joint and
transmission components.
Linkage joints must accommodate constraint forces and moments between
the links. Greater stiffness is achieved with rolling bearings in backlash-free
arrangements. O-arrangements with adjusted clearance or preload are preferred
for joints subjected to large tilting moments due to the increased “virtual” bearing
distance (cf. Section 3.2). Another major concern is the weight of a joint. An
optimal tradeoff is often achieved by using thin-section bearings with reduced
cross sections but larger bore diameters and greater number of rolling elements.
The issues related to an optimal selection and design of actuators and drive
mechanisms are addressed in Section 2.5.2.
In the following, effects of the structure-varying kinematic chain on the leg
joints are analyzed using a simplified 2-D model. For the sake of simplicity it
is assumed that the deflection of the leg endpoint due to link elasticity is small
compared to joint compliance. The leg is approximated as a planar mechanism
of two rigid links, i = 1, 2, each with length li , mass mi and CoM position lic . The
foot is modeled as a lumped mass mfoot attached to the free end of the second link.
Links and foot are connected by flexible joints with stiffnesses khip , kknee and kankle .
During swing and stance phase, the leg rotates about its base joint with constant
angular acceleration, ϕ̈swing and ϕ̈stance , respectively.
For the swing leg, the upper body is assumed fixed in space and the hip joint is
the base joint of the leg (Figure 2.7b). Then deviations of the desired foot motion
are mainly due to hip and knee joint compliances, while the effects of ankle joint
elasticity are less emphasized. By analogy with the above considerations of the
analytical bending beam, the hip joint stiffness must be the highest, followed by
the knee and ankle joint stiffnesses: khip > kknee kankle .
Conversely, the stance leg contacts ground with its free end (i. e., the viscoelastic
layer at the foot) to support the upper body, as illustrated in Figure 2.7c. Without
loss of validity the foot-ground contact is considered as a passive pivot with stiff-
ness kcontact , forming the “base joint” during stance phase. Contrary to the swing
leg, the foot-ground contact and ankle joint must show the highest stiffnesses in
order to minimize upper body deviations, while the knee and hip joint stiffnesses
are less critical: kcontact ≥ kankle > kknee > khip . However, contact stiffness is rated to
effectively attenuate impacts (cf. Section 2.3) and ensure performance of ankle
joint torque control, and distinctly lower than the stiffnesses of the hip joint and
other leg components: kcontact khip . Indeed, upper body deviations during stance
are dominated by compliance in the foot-ground contact.
As with the link structures, alternating leg topologies during the gait cycle yield
conflicting requirements on the leg joints. However, since the highest accelerations
occur during swing phase, the hip joint must evidently show the highest torsional
stiffness. Though torsional stiffness of the the knee and ankle joints can be
distinctly lower, yet large enough to ensure good control performance and high
bandwidth of the torque controller.
Design Approach
In summary, the leg structure must meet the following requirements to achieve
high stiffness and minimal deflection from the desired trajectories:
– Generally high stiffness of link structures in bending and torsion
– High stiffness of the thigh link. The shank link should be designed extremely
lightweight with less strict requirements to stiffness.
– Thin-walled cross sections with high second moment of area. Additionally,
links should taper in cross section and/or wall thickness to reduce associated
gravity and inertial loading (see also Section 3.2).
– High torsional stiffness of the hip joint, followed by the knee and ankle joints
– Joint drive mechanisms and their auxiliary transmission components de-
signed for high acceleration capability and sufficient bandwidth, see Sec-
tion 2.5.2
– Linkage joints in backlash-free arrangements, preferably O-arrangements
with adjusted clearance or preload
– Very lightweight foot design, see Section 2.3
2.4.2 Total Center of Mass Height

Unlike humans, the largest portion of a biped robot’s mass resides in the legs
since the motors and gears account for approximately a third of the overall mass
[46]. Hence the total CoM height is typically at a level with the hip joint or even
below. During single support phase, the dominant dynamics of a biped robot can
be described by an inverted pendulum. Assuming that the robot mass is lumped
in the CoM which is constrained to move at constant height, one gets a set of
two decoupled differential equations for sagittal and lateral CoM motion. This
approach developed by Kajita et al. [88] is known as the “Three-Dimensional
Linear Inverted Pendulum Mode” (3D-LIPM). It illustrates the influence of total
CoM height on the lateral swing of the upper body during walking. One possible
solution for CoM lateral motion are pieces of hyperbolic curves:
r !
g
∆yCoM ∝ cosh T . (2.3)
zCoM s
According to this approximation, the CoM lateral swing ∆yCoM decreases with
higher CoM positions zCoM for a given single support period Ts . Lower lateral
swing augments the stability of the robot at higher walking speeds because the
angular momentum about the fore-aft axis is reduced. Maximizing the height of
the robot’s CoM is therefore a basic design guideline.
Moreover, achieving a human-like mass distribution by concentrating the mass
in the upper part of the robot would be desirable. As pointed out in Section 2.6,
most servo actuators typically are in the robot’s legs. Considering the resultant
mass distribution, it seems unlikely to shift the total CoM above the hip joint
without adding masses to the upper body, cf. Figure 2.6b. In particular, arranging
parts of the leg drive mechanisms in the upper body is very difficult. This approach
normally leads to very complex drive mechanisms with various transmission
components, spanning several links and multi-DoF joints. Moreover, the longer
transmission distance increases the motor-side inertia and may introduce backlash.
In summary, the practical influence on CoM height is limited. Shifting the total
CoM as close as possible to the hip joint is nevertheless an important design
guideline, provided that the robot’s performance is not affected by complicated
drive mechanisms.
2.4.3 Low-inertia Design of the Leg Links

Higher speeds of locomotion normally yield large step lengths at high step fre-
quencies. Conversely, if the robot moves autonomously in realistic scenarios,
sudden and arbitrary changes of the walking speed and direction may occur. Both
tasks benefit from a greater acceleration capability of the locomotor system which
is achieved by a low-inertia design of the leg links. As illustrated in Figure 2.6c, the
mass distribution in the leg links directly affects a variety of robot characteristics,
including the CoM position, the resultant inertia of the legs, and the elastody-
namic behavior of the drivetrains. Therefore, the judicious design of the leg links
and their drive mechanisms is important to achieve (a) high dynamic performance
of the overall system and (b) support the performance and robustness of balance
control.
Regarding the robot’s agility and speed of locomotion, a high dynamic perfor-
mance of the locomotor system is one of the most fundamental design objectives.
The naïve approach is to simply design the servo actuators for high torque and
speed bandwidths. But drive power must be increased very carefully since more
powerful actuators are normally heavier and may in turn easily detract from the
robot’s dynamic behavior. High-performance servo actuators are nevertheless
appropriate, yet with limited benefit if no accompanying measures are taken. In
this light, it is proposed to reduce the inertia and gravity loading of the individual
leg links to improve the acceleration capability of the leg actuators. This is basically
achieved by minimizing the masses of the highly accelerated leg links. In addition,
arranging heavier component parts, in particular the servo motors and gearboxes,
close to the articulated joints of the individual links further contributes to im-
proved acceleration behavior of the locomotor system. Taking this approach to its
logical conclusion, systematically minimizing the resultant inertia of the entire
2.5 Joint Mechanisms 37
leg apparatus and nonlinear effects due to varying inertia can further improve the
acceleration capability of the entire leg. Ideally, the individual link inertias should
decrease from the base to the end link. However, in many conventional biped
robot designs the actuator of a leg link is located at or close to the articulated
joint of the link. Hence the entire leg apparatus becomes heavy and bulky which
detracts from the desired dynamic performance. It is therefore proposed to locate
heavier component parts not only close to the articulated joint of an individual
link but near the pivot point of the hip joint, whenever possible.
Note that also the elastodynamic behavior of the drivetrains benefits from lower
link inertias since the load inertia enters the calculation of the natural frequency
and antiresonance of a drive mechanism, Eq. (3.17) on page 89. The problem is
becomes evident in the hip joint: the flexion/extension drive is subjected to high
torque loading and alternating speeds with high accelerations. Reducing the load-
side inertia increases the antiresonance frequency, along with the capability of
the actuator to compensate disturbances and respond to desired actions. Further
details are discussed in Footnote 6.
Finally, the performance and robustness of balance control can benefit from
a sophisticated leg design. Balance control and real-time trajectory generation
are currently only feasible by using simplified models. Most of them rely on an
“inverted pendulum model” approach, treating the robot as a single point mass.
The modeling error is mainly due to swing leg dynamics and leads to oscillations
of the upper body. Although more exact models, e. g., [17], may compensate
for some of these errors, the stability margin available for balance control is
narrowed by poor leg dynamics if a certain amount of stability reserves accounts
for compensating modeling errors. Conversely, greater stability reserves are
available for actual disturbances of the robot if the locomotor system is optimized
for low disturbances by leg angular momentum and other nonlinear effects.
2.5 Joint Mechanisms

The mechatronic servo systems are among the core units of a walking robot
and composed of four major components: (a) actuators, (b) drive mechanisms,
(c) state sensors, (d) motor electronics (servo controller and drive). In general,
operating performance and suitability mainly rely on how well these components
are matched. Hence they must be considered simultaneously, along with the
kinematic structure of the drive mechanism.
Servo systems for walking robots must cover a large angular velocity range and
have high overload capability to cope with task uncertainty. Since joint positions
are often held for a certain time, it is desirable to have low-wear when the drives
are in reversing operation around zero speed. General requirements on servo
systems also apply, including high efficiency and good behavior in steady-state
and transient operation. Good steady-state operational behavior is achieved by
designing the mechanical hardware for linear system behavior whenever possible.
Incorporating transmission components which are free from backlash and stick-
slip improves positioning accuracy and enables jerk-free acceleration even at lower
speeds. The transient operational behavior is improved by selecting actuators

with small time constants and sufficient torque capacity.
This section focuses on the actuators and transmission components which are
determined by the loads and the kinematic and mechanical structures of the
joints. After discussing the relevant state of the art, the performance and effi-
ciency of different actuator types are analyzed in Section 2.5.1. Suitable precision
gears with high reduction ratios and auxiliary transmission components are intro-
duced in Section 2.5.2. The state sensors and motor electronics are discussed in
Sections 2.7.2 and 2.8.3, respectively.
Related Work
The prevalent actuation principle of current fully actuated humanoids is the
combination of electric motors and reduction gears. Other solutions like fluidic
actuators (hydraulic or pneumatic) are rarely used. Gienger [46] showed that
servo actuators make almost 31 % of a biped robot’s mass: 22.7 % account for
the electric motors and another 7.9 % for the reduction gears. Although early
prototypes of Honda’s biped robots employed ballscrew drives [65], the combi-
nation of DC brush motors and Harmonic Drive gears is the de-facto standard
for fully actuated humanoid robots, for example [95, 100, 126, 130]. The joints
of Asimo [65] and ETL-Humanoid [122] are driven by both DC brush motors and
DC brushless motors2 through Harmonic Drive gears. Johnnie is actuated by DC
brush motors and Harmonic Drive gears but the ankle joints are driven by parallel
mechanisms with ballscrews [47]. The robot BIP 2000 [35] is equipped with brush-
less motors that drive the joints through Harmonic Drive gears or planetary roller
screws. The robots SDR-4X and QRIO are actuated by “intelligent servo actuators,”
comprising brushless motor, planetary gear and electronics. The units are charac-
terized by high backdrivability and seem to be optimized for mass production [40,
82]. A very interesting concept is realized in DLR’s torque-controlled seven-DoF
robotic arm LWR-III [67]. The joints are driven by very efficient brushless motors
called “RoboDrive” through Harmonic Drive gears. These motors are optimized
for the demands of robotic applications and recently became commercially avail-
able. Only few humanoid robots are actuated by fluidic actuators, for example,
hydraulic cylinders [23], or pneumatic muscles [191]. Kaminaga et al. [91] pro-
pose a highly backdrivable electro-hydraulic knee actuator with low mechanical
impedance for better interaction with human beings.
2.5.1 Actuators
Although the literature survey in the previous section suggests certain “conver-
gence” towards the combination of electric motors and Harmonic Drive gears, it is
worth considering other actuation principles for humanoid robots [81]:
Pneumatic actuators have a rather poor power density and are compliant due
to air compressibility which makes them unsuitable for position-controlled
servo applications.
2 See the clarification of brushless motor terminology in Footnote 3 on the facing page.
Hydraulic actuators have an unmatched power density and a high-dynamic be-

havior. However, the hydraulic power unit and the auxiliary system makes
them rather complex and bulky. Provided that the entire hydraulic system
must be integrated on the robot in order to maintain mobility, the power-to-
weight ratio of these actuators decreases rapidly.
The support system of an electric motor on the other hand, consists of the power
electronics and a controller unit and is normally by far less bulky. Therefore, elec-
tric motors are chosen to actuate the robot. Brushed DC motors and electronically
commutated motors are most suitable for servo applications requiring high power
density:
Brushed DC motors (BDCM) The windings of a BDCM are attached to the rotor,
the permanent magnets to the stator. Mechanical commutation allows the
control of motor speed by applying a pulse-width modulated voltage. Sev-
eral designs with optimized speed and positioning performance are available
for servo applications: the bar-armature, disc-armature and bell-type arma-
ture motors. The latter two have an ironless rotor and a winding molded
with epoxy in the shape of a disc or tube. The advantages are low inertia, no
iron losses, and smooth output with low torque ripple. However, duty-cycle
limitations arise from the low thermal capacity and limited thermal paths to
the motor housing. The mechanical commutation principally limits the max-
imum speed and output of a BDCM and severely narrows the permissible
peak torque.
Electronically commutated motors (EC motors) Unlike BDCM, the rotor carries
permanent magnets with a certain number of pole pairs. The stator assembly
consists of stacked steel laminations holding the phase windings. Hence
EC motors are synchronous motors excited by permanent magnets. EC
motors exhibit iron losses and cogging torque, the latter can be minimized
by a skewed stator and larger pole count. Though multiphase motors are
advantageous for some applications that require low-speed operation with
high torque as well as high-speed operation, the typical winding arrange-
ment is a three-phase configuration in either star or delta connection. EC
Motors can be distinguished by the shape of the back-electromotive force
(back-EMF): motors with trapezoidal and sinusoidal back-EMF are referred
to as Brushless DC Motors (BLDCM) and AC Permanent Magnet Synchronous
Motors (PMSM), respectively3 . For constant torque output trapezoidal and
sinusoidal currents must be applied to the BLDCM and PMSM, respectively.
The decided advantages of EC motors over BDCM are the better robustness and
reliability, significantly higher power density and acceleration capability due to
the large torque and speed bandwidths. The continuous stall torque rating of EC
motors is significantly higher due to excellent cooling properties and low thermal
resistance to the environment. Moreover, electronic commutation permits larger
3 Several ill-defined terms are used and mixed up in the literature and technical documents
on EC motors: since “AC servo motors” are often referred to as induction motors, the term
“AC Permanent Magnet Synchronous Motor” (PMSM) is used for EC motors with sinusoidal
back-EMF. EC motors with trapezoidal back-EMF are referred to as “Brushless DC Motors”
(BLDCM).
peak torques for longer intervals which is of particular importance when the
motors are in reversing operation around zero speed. Although EC motors are
simpler in design than BDCM, they require more complex controllers and motor
electronics. But the integration level of current power and control electronics
allows the cost-effective and extremely compact design of the necessary electronics,
making EC motors the de-facto standard for high-performance servo motors in
industrial applications.
Comparing the different types of EC motors, BLDCM are apparently easier in
control and rotor position sensing: three Hall sensors provide six different switch
combinations per (electrical) revolution. The rotating field for a PMSM is usually
generated by field-oriented control [149, 166] and requires high-resolution rotor
position feedback (encoder or resolver). Both motor types are comparable in small-
signal and large-signal dynamic behavior. At higher speeds the commutation
torque ripples of a BLDCM are damped out effectively by rotor inertia but severely
affect low-speed performance. In particular, the accuracy and repeatability of
position-servo performance deteriorates [166]. PMSM on the other hand, have the
advantage of lower acoustic noise and torque ripple and are therefore preferred in
high-performance speed and position-servo applications.
Servo Motor Selection Rationale
Volume and size of the motors are critical factors in high-performance applications
such as robotics. In order to realize the desired motion patterns, the motors
must exhibit high power density and provide sufficiently large torque and speed
bandwidths. Before actually dimensioning the servo motors (see Section 3.3.2), a
motor family is selected by comparing the technical data using different indices.
Several performance measures for robotic servo systems were proposed. For
example, the “joint quality measure” by Hirzinger et al. [66] takes into account
peak torque, maximum velocity and mass of the servo motor and transmission
components. Gienger [46] proposes the “specific dynamic continuous torque”
which takes into account motor mass, motor and gear inertia, gear ratio and gear
friction. Both indices are based on parameters that can only be determined from
the final design. Thus, they are of limited use for component selection in an early
design stage, where only catalog data are available. For the pre-selection of a
motor family, benchmarks based on publicly available data are necessary. Again,
various selection criteria exist [103], and the most relevant ones are:
1. Power density
2. Acceleration capability
3. Efficiency and overall performance
Power density P /mm [W/kg] cannot be determined reliably if the motor phase
voltage is not known yet. Therefore, the ratio of continuous stall torque and motor
mass τc /mm [Nm/kg], referred to as “specific continuous torque,” is proposed to
indicate power density. The motor’s acceleration capability can be seen from the
peak-torque-to-inertia ratio τp /Jm [rad/s2 ]. The so-called “dynamic factor” indicates
motor acceleration at maximum torque without external load. Usually the inverter
is not rated to deliver the current corresponding to the peak torque τp , hence the
continuous-torque-to-inertia ratio τc /Jm [rad/s2 ] is the more reliable figure. Though

√
almost not present in the scientific literature, the motor constant Km [Nm/ W] is
very useful for comparing the efficiency and overall performance of permanent
magnet motors. Assuming a motor at stall, the current I produces a certain peak
torque τp but no output power. Then ohmic losses are the only significant power
dissipation mechanism and Km is defined as the ratio of τp to the square root of
input power, Ps = I 2 R. Hence Km is also the ratio of the torque constant Kτ to the
winding resistance R:
τp K
Km = √ = √ τ . (2.4)
Ps R
Although originally defined with respect to larger direct drive motors operating
at stall or at very low speeds, the motor constant gives an idea of motor efficiency
and is useful for comparing different motors.
Note that all proposed indices are generally winding-independent for a given
size motor. The winding is determined by the phase voltage and current require-
ments, whereas the motor size (diameter and stack length) and power (torque and
speed) are determined by the load and performance requirements.
PMSM/BDCM Market Survey
Figure 2.8 compares servo motors from different manufacturers4 . Figure 2.8(a) and
(b) show the specific continuous torque τc /mm as a measure of power density and
the continuous-torque-to-inertia ratio τc /Jm indicating the acceleration capability,
respectively. Figure 2.8(c) shows the motor constant Km which relates to the
motor efficiency. For the best power-to-weight ratio, only motors using rare-earth
magnetic material (NdFeB, SmCo) are considered. All motors have an inner rotor
design, that is, the rotor is designed inside the stator. This design is preferred in
incremental motion applications where low inertia and fast response are required.
Note that the analysis considers both frameless and housed motors. Frameless
motors (or kit motors) are delivered as individual components (stator, rotor, feed-
back sensor) and allow for very compact and lightweight actuator designs. Motor
housing and cooling can be tailored to the application, enabling the tight integra-
tion into the mechanical structure. The motor shaft and bearing arrangement are
designed in the same manner, matching the precise needs of the application. In
many cases the direct coupling of frameless motor and reduction gear is feasible
and eliminates auxiliary transmission components such as belts, pulleys, auxiliary
gears and couplings. Hence the drivetrain is less complex, free from backlash
and slip, and the resultant stiffness and bandwidth increase. The housings of
commercial motors, on the other hand, are industrially ruggedized and typically
made from steel. Also, the rotor inertia of a housed motor includes the shaft and
bearings, whereas the inertia of a frameless motor only considers the permanent
4 The analysis compares commercially available BDCM and PMSM from Aerotech [1, 2], Axsys
[8, 9], Kollmorgen [28], Maxon [115], Parker Bayside [141], RoboDrive [156], and Wittenstein
[202]
(a) Specific continuous torque (b) Continuous-torque-to-inertia ratio

5 1.8 b
1.6
b
b
4
b
b
b
1.4 b
τc /Jm [106 rad/s2 ]

b
1.2 b
τc /mm [Nm/kg]
b b
3
b b b
b
b
b
b b b
b
b b
1.0 bb
b
b b
b bb b b
0.8
b
b b b b
b
b bb
2 b
b b r Lola b
b b
b
b
Lola
b
b
b
b b
b b
b
b b
r b
b
b
b
r b
b b b b b
0.6 b
b b
r r bbb b
bb b b b b
b b b b b
bb
0.4
b b brb b b b
b bb b bb b
1 b b b bb b b
b bb
rr
b r r
r b b
r
b
b
b b
b
b
b b
b
b
b
b
b rr r b b b b
b b b b b b b
bb r r r r r b b bb
0.2 b b b b b b b
Johnnie
r b
bb b b b rr r r bb b b b b
b
b b r r r r b rb b b r r b b b
Johnnie
r r r r b b
br r b r r rrr r b
rr r r r b r r b b
rb b b r b b b b b b b b b b b b b
rbrb bbr r r r
r r
b b b b
r r r rb r b b b r r b b rrr r
r r b
0 0 r b
b r r b r b r rr b r r br r r r r b b r r r b b b b
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
τc [Nm] τc [Nm]
(c) Motor constant

1.0
b
0.8
r
Aerotech (BDCM, housed)
b
Aerotech (PMSM, frameless)
r
Axsys (BDCM, frameless)
Km [Nm/ W]
b b
0.6 r
b b
b
Axsys (PMSM, frameless)
√
b
r b b
r
Maxon (BDCM, housed)
b b
0.4
b
b
b
b
Kollmorgen (PMSM, frameless)
r
Parker Bayside (PMSM, frameless)
b
b bb b b
b
b
r b b
b b
b
b b
b
b b
b
b b
RoboDrive (PMSM, frameless)
r rrr r b b
0.2 r b b b
b
r b b
Wittenstein (PMSM, housed)
Lola
b
b rr b b b
b rb
r bb b b b b
r
b rr b b b b r b
b
Johnnie
b
bb r br r b r r b b b b b
b r r r r b b br b b
b r r r r b b b b b b b b bb b b
r bb
r r bb bb b b b
0 r b r b b bb b b
0 1 2 3 4 5 6 7 8
τc [Nm]
Figure 2.8: Power density, acceleration capability and efficiency of commercial BDCM
and PMSM, represented by the (a) specific continuous torque; (b) continuous-torque-
to-inertia ratio; (c) motor constant.
magnets. Thus, the comparison between frameless and housed motors using mass-
relateed benchmarks is limited. Nevertheless, servo actuators using frameless
motors are supposed to have a significantly higher power-to-weight ratio and
acceleration capability than actuators based on standard framed motors with the
same power rating.
Obviously, PMSM are clearly superior to BDCM both in specific continuous
torque and continuous-torque-to-inertia ratio. RoboDrive [156] offers frameless
PMSM in a segmented lamination design with a high pole count that are opti-
mized for robotic applications. The motors are characterized by a high torque
capacity at low power consumption and losses. Due to the high pole count, these
motors are optimized for high torque delivery and efficiency at low and medium
speeds. However, motors with high torque and speed bandwidths are required
for a humanoid walking robot which is targeted to achieve human-like speeds
of locomotion. Frameless PMSM from Parker Bayside [141], on the other hand,
show excellent dynamic capabilities at a power density comparable to RoboDrive.

Ripple and cogging torque of these motors are minimized by a skewed stator
design. Moreover, the large variety of available winding configurations offers great
flexibility in matching the motor to the application. Therefore, this motor series is
chosen for all joint of the robot, except for the camera head where only very low
torque capacity is required.
Compared to the robot Johnnie, the power density and acceleration capability
rise substantially as indicated by the shaded areas in Figure 2.8. The hip and knee
joints are driven by two Maxon motors RE-40 each [115]. As a straightforward
measure for enhancing the performance, Gienger [46] proposes to install a third
RE-40 motor in parallel. The power density remains √ constant
√ in this case but,
according to Eq. (2.4), Km only increases by factor 2 and 3, respectively, if two
or three motors are used in parallel. Conversely, the specific continuous torque
of Lola’s servo motors increases by factor 3.7 for low-performance up to 6.8 for
high-performance motors. The acceleration capability increases by factor 3.7 to
12.6 at comparable efficiency.
2.5.2 Drive Mechanisms

The drive mechanism interfaces the motor to the robot link through the joint. It is
composed of the main transmission (typically a reduction gear) which interfaces
to the joint, and—depending on the structure of the mechanism—additional
components for power transmission such as coupler links, belts, or angular gears.
The proper placement, sizing and design of the drive mechanisms sets the stiffness,
mass, and overall operational performance of the robot [163]. Obviously, all
transmission components have to provide the desired torque rating and keep
certain overloading capacity, combined with good dynamic performance and
sufficient torsional stiffness. Additionally, they must have compact dimensions,
low mass and low inertia. Another important factor to achieve high efficiency
and reduce the risk of damage in an emergency, is backdrivability5 of the drive
mechanism. In the following, the main transmissions and auxiliary components
for power transmission are introduced.
Direct Drives
The kinematically simplest actuation mechanism interfaces an EC motor with
high-torque and low-speed characteristics between the links without any reduction
gear. Since mechanical gearing is completely eliminated, drive stiffness is high,
friction is significantly reduced, and mechanical backlash and hysteresis are
removed. Direct drives are fully backdrivable which would be particularly relevant
for the sagittal plane joints: similar to human walking [142], the knee and hip
actuators would accelerate the links at selected intervals of the gait cycle, while
5 Backdrivability is the ability to move the output of the drive mechanism throughout the
workspace without opposition. It is usually characterized by the backdriving friction of the
mechanism. Backdrivability is usually reduced by certain elements, such as gears and fric-
tion in the motors and drive mechanisms.
3
2
1 Wave Generator 2 Flexspline 3 Circular Spline
Figure 2.9: Working principle of the Harmonic Drive gear. The Circular Spline is fixed
to the environment and the Wave Generator is driven by an actuator. The arrows
indicate the relative position of the Wave Generator and Flexspline.
limb advancement would occur virtually passive. Despite excellent dynamic

properties, direct drive motors are heavier and have larger physical dimensions,
compared to conventional motor-gear units and require more accurate position
sensors and controllers. Therefore, motor-gear units are currently the drive of
choice for humanoid robots.
Harmonic Drive Gears

As stated above, Harmonic Drive (HD) gears are the de-facto standard for full-size
humanoid robots. The principle of operation is described in detail in [59] and it
is summarized here for clarity. The gear consists of three principal components
shown in Figure 2.9. The elliptical Wave Generator (1) with ball bearing is usually
driven by an electric motor. The Flexspline (2) is a flexible steel cylinder with
external toothing, and the Circular Spline (3) a steel ring with internal toothing.
The Flexspline has two fewer teeth than the Circular Spline. It is held in an
elliptical shape by the Wave Generator, so that its teeth engage with the Circular
Spline across the major axis of the ellipse. If the Wave Generator rotates, the zone
of tooth engagement travels with the major elliptical axis, and each clockwise
turn of the Wave Generator moves the Flexspline two teeth anti-clockwise with
respect to the Circular Spline. The advantages of HD gears are well known and
include excellent positioning accuracy and repeatability, high torque capacity
and superior torque-to-weight ratio, compared to other gear types. HD gears
have high single-stage reduction ratios (30 ≤ N ≤ 160), may operate with zero
backlash6 , and exhibit high torsional stiffness over the full speed range. Although
they can achieve efficiencies of up to 85 %, the frictional behavior is complex and
depends for instance on manufacturing and assembly tolerances [177]. HD units
are reversible and backdrivable. However, efficiency decreases significantly when
in backdriving: a comparison of catalog data shows that the “no-load backdriving
torque” is approximately 23 % higher than the “no-load starting torque” required
to commence rotation of the Wave Generator with no load applied to the output.
6 The gear teeth of HD gears have no backlash. But there are hysteresis losses (typically
<3 arc min) and so-called “lost motion” (typically <1 arc min), resulting from the torsional
stiffness in the low-torque region [59].
1 Wave Generator
2 Flexspline
3 Circular Spline
Figure 2.10: Harmonic Drive component set in a customized lightweight version. The
Circular Spline has a T-shaped cross section, the Wave Generator plug is a solid body
made from aluminum without Oldham coupling.
Due to the many advantages, HD gears are employed as speed reducers in

most joints. As shown in Figure 2.10 the customized lightweight versions have
a T-shaped Circular Spline to reduce weight and inertia while maintaining high
radial stiffness. The solid Wave Generator plug has an optimized geometry and
omits the Oldham coupling of the standard versions. To further reduce weight
and inertia, special aluminum with almost the same thermal expansion coefficient
as steel is used for the Wave Generator plug of some gears. Compared to other
experimental lightweight solutions [174], these modifications are an economically
justifiable tradeoff between weight and load rating. During the robot development,
the new CPL Series Lightweight Component Sets [57] became available as off-the-
shelf products, showing basically the same design features as the custom-made
lightweight versions used in the robot. For joints that are subjected to smaller
loads (e. g., the camera head), off-the-shelf gearboxes are the best choice with
regard to weight and compactness: the CSF Mini Series Gearboxes [58] are very
compact, standardized gearboxes with integrated rigid four-point contact output
bearing. The RSF supermini servo actuators [60] combine gearbox, AC servo motor
and encoder in a very compact package.
Ballscrews and Planetary Roller Screws

The drive mechanisms of knee and ankle joint (cf. Section 2.6 on pages 53–56) re-
quire the transformation of the motor’s rotary motion into translational movement.
Ballscrews and planetary roller screws are perfectly suited for such applications.
They are characterized by high efficiency, smooth and silent operation without
jerk, and no stick-slip:
Ballscrews use balls as rolling members to transfer the load between the screw
and nut. These machine elements are widely used in mechanical engineering
1
1 Screw shaft
3 2 Nut
3 Roller
4 4 Carrier
Figure 2.11: Mechanical design of a planetary roller screw (adapted from [78])
so that a large variety of types is available, e. g., [173], including special and
application-specific designs (actuated nut systems, large-lead screws, etc.).
The permissible rotational speed is limited by the ball recirculation system
in the nut housing. The static and dynamic load rating are determined by
the number of contacts and the contact areas between the balls and the nut
or screw, respectively. As the number of contact points is limited by the
ball size and other design factors, the obvious way of increasing the load
capacity is to select a screw with larger diameter. For joints that must cover
a wide speed range and are subjected to high loads (such as the knee joints),
however, shaft inertia and weight increase significantly. A very expensive
measure is to employ customized hollow-bored screw shafts which have
reduced inertia and weight and higher critical rotating speed but lower axial
stiffness and maximum permissible compressive load.
Planetary roller screws have grooved rollers as rolling members and are less
common than ballscrews. The components are shown in Figure 2.11. The
screw shaft (1) has a multi-threaded profile. The nut (2) has an internal
thread which is complementary to that of the screw shaft. The rollers (3)
have a single-threaded profile and are constantly engaged with the screw
(1) and the nut (2). In addition, rollers and nut are geared together to avoid
slip and the rollers are supported by carrier (4) to insure equal spacing.
The operating mechanism of a planetary roller screw is analogous to the
motion of a planetary gear, and one revolution of the screw shaft results in
an axial displacement on the nut corresponding to its lead [78]. The number
of contact points is significantly higher than at similar-sized ballscrews
and goes along with considerably higher static and dynamic load ratings.
Moreover, the large number of contact points results in greater axial stiffness
and better tolerance against shock loading. Since there is no recirculation
system necessary for the rolling members, the rotational speed is limited only
by the centrifugal forces in the nut. Planetary roller screws can therefore
be operated at higher speeds and can withstand higher acceleration and
deceleration rates.
For minimal inertia it is advisable to actuate the lead screw shaft and move the
nut along the shaft. The rolling contact makes ballscrews and planetary roller
Table 2.2: Comparison chart of Harmonic Drive gears, ballscrews and plane-
tary roller screws
Criterion Harmonic Drive Ballscrew Planetary roller screw
Load ratings ++ + ++
Lifetime ++ ◦ ++
Output speed + ++ +
Positioning accuracy ++ ++ ++
Backlash No∗ No† No†
Backdrivability ◦ + +
Stiffness + +† ++†
Shock loads ◦ ◦ ++
Space requirements ++ + ++
Available range of reduction ratios + ++ +
Efficiency at nominal torque ≥ 85 % ≥ 90 % ≥ 90 %
Backdriving efficiency ◦ ++‡ ++‡
Power density ++ + ++
∗ There are hysteresis losses and so-called “lost motion” in the low-torque region.
† Pre-loaded nuts eliminate backlash and increase stiffness.
‡ The backdriving efficiency depends on the pre-load, screw diameter and pitch.
screws backdrivable under almost all circumstances. The frictional behavior

of ballscrews and planetary roller screws is similar: efficiency depends on the
working conditions (e. g., load, speed, lubrication, preload) and is typically above
90 %. The “direct efficiency” which defines the torque required to transform the
rotation of one member into the translation of the other, and “indirect efficiency”
which relates to the axial load required to transform the translation of one member
into the rotation of the other one, differ by less than 0.5 %. Hence the reversal of
movement is much more efficient than at Harmonic Drive gears. The nuts can
be preloaded to eliminate backlash and/or to increase axial stiffness, however,
internally preloaded screws exhibit higher friction.
The key characteristics of ballscrews and planetary roller screws are compared
to Harmonic Drive gears in Table 2.2. In conclusion, planetary roller screws are
employed as main transmissions of the knee and ankle joint drives (cf. Section 2.6),
whose design is particularly constrained by space and weight. In addition, their
robustness against shock loads makes planetary roller screws particularly suitable
for the robot’s legs. The higher load rating allow to use smaller screws opposite
to ballscrews which finally yields additional weight savings. Due to the reduced
diameter, the buckling at compression loading and the critical rotating speed of
the screw shaft have to be considered in the dimensioning of the drivetrain and
selection of the bearing arrangement.
Other Precision Gear Types with High Reduction Ratios

In addition, two other high-reduction gear types for position servo applications
shall be mentioned. However, these gear types are only relevant for higher power
ranges because lightweight design are barely feasible:
Planetary gears: through the uniform load sharing of several gears, planetary
gears are a very compact solution with high torque ratings and low backlash
(typically in the range of 4 arc min or even ≤1 arc min for position servo
applications) at efficiencies up to 97 %. Multi-stage planetary gear trains
can achieve reduction ratios well over 100. Despite these advantageous
characteristics, planetary gears are not suitable as main reduction gears
because of an unfavorable power-to-weight ratio. On the other hand, the
overall torque-to-weight ratio of an actuator can be improved by integrating
a planetary pre-stage within the Wave Generator of a Harmonic Drive gear
[174]. Such units provide high reduction ratios in a compact space envelope,
enabling the use of high-speed motors with a high power density.
Cycloidal gear reducers are widely used in the joints of industrial robots. They
are characterized by high torsional stiffness, compact measurements and a
high efficiency resulting from the rolling contact of all major load transmit-
ting components.
Auxiliary Transmissions
Actuator placement is a tradeoff between weight, link inertia, compliance and
complexity of the drive mechanisms. The early actuation concept of the robot
aimed at the direct coupling of electric motor and gear unit without any couplings
and other auxiliary transmission components in between. Besides reduced weight,
such drivetrains are free from backlash and slip and have higher stiffness and
system bandwidth. On the other hand, the considerations made in Section 2.4
suggest that actuator relocation by means of auxiliary transmissions is essential
to optimize the mass distribution of the legs for high acceleration capability. The
longer transmission distance reduces inertia and gravity loading but can introduce
compliance and backlash effects. System stiffness and transmission accuracy may
be impaired and oscillations of the servo drivetrain may occur. In summary, the
positive effects of actuator relocation on global system dynamics are much more
significant than the weight savings and greater drivetrain stiffness in the case of
the direct coupling of motor and gear unit. Possible transmission elements for
actuator relocation are band drives, cable drives, synchronous drive belts, bevel
gears and coupler links:
Band drives (e. g., steel, titanium, or aramid cord belt) are generally very stiff
and accurate transmission members. They are widely used in industrial ap-
plications for the transmission of limited-travel rotational and translational
movements, e. g., in SCARA-type robots. Because of the high stiffness and
load rating, band drives are usually implemented at the output of a torque
motor or conventional motor-gear unit, allowing to separate the entire servo
actuator from the joint axis. Using a band drive, it is difficult to span one or
more joint axes.
Cable drives are similar to band drives but steel wire ropes or bowden cables are
used as transmission members. Cable transmissions are used as tendons in
the WAM Arm built by Barrett Technology7 and some humanoids [7, 188],
allowing for a low-inertia design and the actuation of multi-DoF joints in a
7 Available at: http://www.barrett.com/
differential arrangement. Although it is possible to span one or more joint

axes, the stiffness of cable drives is considerably lower than of band drives.
Also, it is difficult to achieve a backdrivable wire-driven joint which is free
from backlash.
Synchronous drive belts, or timing belts8 , transmit torque and motion from a
driving to a driven pulley. They are free from backlash and slip and provide
an accurate way for modest speed reductions with an efficiency of 95 % or
above. Belt drives are mostly used for power transmission between the motor
shaft and the input of a reduction gear, enabling the spatial separation be-
tween the electric motor and main transmission, as well as the space-saving
integration of the motors into the robot structure. Synchronous belt drives
add considerable damping to the system which allows higher gain settings
of the motor controller. A disadvantage of belt drives is the added inertia
of the belt and pulleys, the latter can be minimized by modifying standard
pulleys made from aluminum or plastic. Moreover, synchronous belts add
compliance to the drivetrain which can impair transmission accuracy and
cause oscillations, if not designed and installed properly [24]. The resultant
stiffness of a belt drive is composed of the belt stiffness of the entering (tight)
and leaving (slack) sides of the driven pulley, and the belt teeth and cords
along the arc of tooth meshing. The most important measure to limit com-
pliance effects is to choose belts with high tensile strength and optimized
tooth profile, for example HTD Synchronous Drive Belts [24]. In addition,
belt pre-tension must be set by an adjustable pulley or idler in order to
prevent the belt from sagging and ensure proper tooth meshing under all
circumstances. Proper belt pre-tension also increases the tooth stiffness
along the arc of tooth meshing.
Angular gears: the drive mechanism of the ankle joint requires the redirection of
the axis of rotation from the knee joint axis to the longitudinal direction of
the shank (cf. Section 2.6.3). Spiral-toothed precision bevel gears designed
for servo applications are used [55] because of smooth power transmission,
high load rating at an efficiency ranging from 94 % up to 97 %. Spiral bevel
gears have the benefit of very favorable meshing characteristics (in terms of
high contact ratio) and curved teeth which are considerably more resistant
to distortion than straight or helical teeth.
Coupler links are used in the parallel mechanisms of the ankle joints to connect
the outputs of the linear actuators to the foot segment. Under normal oper-
ating conditions, coupler links are only subjected to normal forces, enabling
a very slender and lightweight design that must only be dimensioned for
buckling. Coupler links normally employ spherical and/or universal joints
with plain bearings at their ends. Because plain bearings normally introduce
backlash which impairs the performance of the joint controller, bearings
with reduced clearance must be chosen.
8 Although the term “timing belt” originates from combustion engine technology, it is widely
used in mechanical engineering. Still, the term “synchronous belt” which originates from
industrial power transmission is used throughout this thesis.
2.6 Kinematic Structure of the Leg Joints

In this section the kinematic structures of the leg joints are developed with special
emphasis on the design objectives devised in Section 2.4. The kinematic structure
of the robot comprises single-DoF revolute joints and multi-DoF joints. Single-DoF
revolute joints, such as the knee and toe joints, can be actuated by either rotary
or linear drives. Multi-DoF joints, for example the hip and ankle joints, can be
actuated either in a serial or parallel configuration.
Rotary versus Linear Actuation of Single-DoF Revolute Joints

The standard actuation principle for single-DoF revolute joints is the rotary drive
where a Harmonic Drive gear that is coaxially with the joint axis interfaces the
motor to the robot link. In this case the output bearing of the reduction gear
coincides with the joint. Depending on the surrounding construction and/or
stiffness requirements, this bearing is designed either as single bearing that can
accommodate loads in all directions (radial, axial, moments) or a bearing arrange-
ment (adjusted or locating/non-locating). Spatial separation between the motor
and gear unit is rather simple by using auxiliary transmissions, e. g., synchronous
drive belts. Moreover, relocation of the whole servo actuator (motor and gear
unit) from the articulated joint is accomplished, for instance, by a parallel-crank
mechanism that interfaces the gear output to the driven link. Rotary drives based
on Harmonic Drive gears are simple in design and control due to the linear rela-
tionship of input and output. All rotary drives of the robot can be built on the
unit construction principle, thus, they share the identical design but differ in size
and output. In conventional leg designs all joints are driven by such “conven-
tional” revolute joints, forming an open kinematic chain of a series of cantilever
beams. While the mechanical design is rather simple, the mass distribution of
such structures is adverse, if the motors are coaxial or co-located with the articu-
lated joints. Moreover, insufficient static and dynamic stiffness of the structure
and transmission components (cf. Sections 3.3.2 and 3.3.3) can severely limit the
dynamic performance of the locomotor system, or at worst destabilize the robot.
It is alternatively proposed to drive a single-DoF revolute joint through a mech-
anism based on a linear actuator. The actuator connects to the driven link for
example by a slider-crank mechanism or acting as a length-variable coupler within
a four-bar mechanism. Again, this approach enables the spatial separation be-
tween the entire drive from the joint to reduce its inertial contribution. The
transfer behavior of such mechanisms is nonlinear which makes joint control more
difficult since the nonlinearities must be taken into account. On the other hand,
the non-uniform transmission ratio can be utilized advantageously by adjusting
the mechanism to the torque-speed characteristics of typical movements of the
robot. For example, the mechanism can be designed such that the transmission
ratio increases where higher joint torques occur, and decreases in ranges of higher
speeds. Moreover, the stiffness of the joint and surrounding structure increases
as the leg link and the linear actuator and the mechanism form a closed kine-
matic chain. However, four-bar mechanisms are more complex in design and
manufacturing, and the several bearings may introduce backlash.
2.6 Kinematic Structure of the Leg Joints 51
Serial versus Parallel Actuation of Multi-DoF Joints

In case of serial actuation the individual joint axes are treated as single-DoF joints,
so that the above considerations apply here, too. Since each axis of rotation is
actuated independently, the kinematics are easy to control. Mechanical end stops
can be applied easily and the unit construction principle of the motor-gear units
simplifies the mechanical design.
Driving a multi-DoF joint in a parallel arrangement often results in a more
uniform load distribution among the actuators, hence the drives can be smaller
in size compared to the serial arrangement, which goes along with additional
weight savings and less power consumption. Parallel kinematics can be actuated
by rotary and/or linear drives which are often situated on the base link. In contrast
to a serial configuration, the moving mass and inertial contribution are reduced
significantly. The stiffness of parallel structures is usually high but depends
on variable lever arms that result from the moving articulation points of the
mechanism. The workspace is to a large extent constrained by the mechanism
itself, i. e., by kinematic singularities and self-collisions. Hence the ratio of the
workspace to the size of the mechanism is sometimes poor. Also, self-collisions
determine the shape of the workspace, which is typically non-rectangular. In
consequence, mechanical end stops are difficult to apply and of limited use.
Rather, the software-based monitoring of the workspace is necessary to avoid
self-collisions. The design, manufacturing and assembly of parallel mechanisms
are more involved due to the higher part count and a large number of bearings. It
is thus more difficult to achieve the same accuracy as with a serial mechanism.
The kinematics of parallel mechanisms are typically described as a set of non-
linear equations. Calculating the manipulated variables for motor control in
real-time is computationally expensive, in particular if no closed-form solution
exists. Therefore, the solution is often obtained off-line and read out from lookup
tables in real-time mode.
2.6.1 Hip Joint Actuation

The three-DoF hip joint is the root joint of the cantilever leg structure. Situated
between the thigh and pelvis, it allows three independent rotations of the thigh.
During locomotion, the hip joints are subjected to high moment loadings that
result from the high impact forces at initial contact and the propulsive forces and
angular momentum of the swing leg. Therefore, a well thought-out design of
the kinematics and mechanical structure are crucial to the dynamic performance
of the legs. In particular, high stiffness of the hip joint structure and bearings
minimizes oscillations of the leg endpoints that may at worst destabilize the robot.
Related Work
All full-size humanoid walking robots incorporate three-DoF hip joints. As an
example, the hip joint of the robot Johnnie is built in a serial configuration. The
internal/external rotation joint is screwed to the pelvis. For better power distri-
bution among the drives this joint is inclined 15° to the vertical axis [46]. It is
followed by the adduction/abduction joint and the flexion/extension joint which
interfaces to the thigh. While the first two joint axes intersect in one point, the
adduction/abduction and flexion/extension axes are offset from one another for
simplified mechanical design and easier integration of the actuators. A different
structure is applied to the robots HRP-2 and HRP-3. The internal/external rotation
axis is situated close to the sagittal plane, resulting in a cantilever structure of the
hip joint. The so-called “cantilevered crotch joint” reduced collisions between the
thighs and allows for walking in confined areas [95, 96]. Nakamura et al. [123]
propose a “double spherical joint” for the robot UT-θ. The spherical hip joints of
both legs have a shared center of rotation in the sagittal plane to fully decouple
the movements of the trunk and legs. To the author’s knowledge, very few robots
have a parallel mechanism applied to at least two hip joint DoFs. Konno et al.
[102] adopt a mechanism to the robot ROBIAN. A parallel mechanism composed
of two linear actuators drives the abduction/adduction and internal/external
rotation joints. The flexion/extension joint is driven independently by a rotary
drive. Hofschulte et al. [70] develop a three-DoF spherical parallel hip joint
for the robot LISA. All motors are situated in the pelvis. A spherical ball bearing
unloads the mechanism from the robot’s mass. The complex mechanism is com-
posed of several spatial links and shows considerable compliance. The three-DoF
hip joint of the robot iCub uses a two-DoF cable differential mechanism in the
flexion/extension and abduction/adduction axes, the internal/external rotation
of the thigh is actuated independently [188]. Finally, Rosheim [158] provides an
extensive survey of three-DoF wrist designs for industrial robots on the basis of
serial and parallel actuation principles, which can serve as a source of inspiration.
Solution Approach
The individual axes of the hip joint show very different loading profiles (see Ap-
pendix D). While the internal/external rotation and adduction/abduction axes are
subjected to moderate torque loadings at small motion arcs, the flexion/extension
axis is among the highest loaded leg joints. These characteristics suggest the use
of a parallel mechanism for the hip joint. However, the large angular workspace
of the flexion/extension axis makes the realization of a reliable three-DoF paral-
lel mechanism very difficult. The problem is reduced by combining a two-DoF
parallel mechanism with a third axis that is driven independently. While the
comparatively small angular ranges of the internal/external rotation and ad-
duction/abduction axes allow the design of a mechanism without considerable
posture-dependent stiffness variations, this architecture does not fully utilize the
advantage of load distribution. The highly loaded flexion/extension axis with is
large motion arc is actuated outside the parallel structure. Conversely, a mech-
anism comprising the adduction/abduction and flexion/extension axes avoids
these shortcomings, however, the large workspace of the flexion/extension axis
makes the realization very difficult. Moreover, the posture-dependent stiffness
may cause unacceptable deflections or structural vibrations at higher loads.
The positive experience of the Johnnie project, on the other hand, suggests
to actuate the hip joint in a serial configuration. The proper kinematic and
mechanical design ensures a high static and dynamic stiffness and can improve
the load distribution among the drives. Three serial drives compose a spherical
joint with all axes intersecting in a single point. The principle axes arrangement
is identical to the robot Johnnie because it avoids singular configurations in the
desired range of motion: (a) internal/external rotation, (b) adduction/abduction,

(c) flexion/extension. Furthermore, the inclination of the rotation joints is kept
for a better power distribution among the drives [46].
In conclusion, serial actuation is preferred for the hip joint because the mechan-
ical design is less complex compared to parallel mechanisms. The development
risk is low, provided that the connecting links and joint bearings are designed for
high stiffness and direct force flux.
2.6.2 Knee Joint Actuation

The knee is the second highly loaded joint in the kinematic chain of the leg.
The goal is to develop an actuation principle which allows to enhance actuator
performance while thigh inertia is kept minimal.
Related Work
The knees of all full-size humanoid robots are designed as one-DoF hinge joints
and mostly actuated by Harmonic Drive gears. Using synchronous belt transmis-
sions, the electric motors are situated either in or close to knee joint axis, e. g., [46,
140] or close to the hip joint, e. g., [94, 127]. A very different concept is realized in
the robot UT-θ: the knee joint allows only unilateral torque transmission from
the knee actuator to the shank. The so-called “backlash clutch” allows the shank
to swing freely, exploiting its passive dynamics [134]. Espiau and Sardain [35]
employ slider-crank mechanisms in the knee joints of the robot BIP 2000. The
planetary roller screw drive is fixed to the thigh and connects to the shank with
a coupler link. A linear bearing is required to keep the roller screw free from
radial loads. A patent application filed by Honda [51] claims another approach
to reduce thigh inertia: the rotary drive (motor-gear unit) for knee actuation is
placed close to the hip joint. The output torque is transmitted to the knee joint by
a parallel-crank mechanism. Another patent application by Honda [181] proposes
a counterbalancing system based on a gas spring, similar to the mechanisms em-
ployed in industrial robots. The gas spring acts in parallel to the knee actuator
and serves to minimize the moments generated about the knee axis when the knee
is in motion and at rest.
Solution Approach
In the following, different kinematic configurations of the knee joint are analyzed
in terms of dynamic properties and simplicity of design. Figure 2.12 depicts
four possible kinematic chains. Although all mechanisms in question have a
planar structure the kinematic analysis is conducted in 3-D, taking into account
manufacturing and mounting tolerances, as well as elastic deformations during
operation. Kutzbach-Gruebler’s equation is used to calculate the degree of
freedom or mobility of a kinematic chain:
j
X
m = 6(n − 1 − j) + fk . (2.5)
k=1
4 7 8
9 10
9
1
1 1 1
11
6
12
12
5 3 3 3
9 8
9
ϕknee
2 2 2 2
(a) Solution K1 (b) Solution K2 (c) Solution K3 (d) Solution K4
1 Thigh 5 Gear unit (b = 1) 9 Spherical joint (b = 3)

2 Shank 6 Belt drive (b = 1) 10 Linear carriage (b = 1)
3 Knee joint (b = 1) 7 Motor-gear unit (b = 1) 11 Linear actuator (b = 2)
4 Electric motor (b = 1) 8 Universal joint (b = 2) 12 Coupler link
Figure 2.12: Kinematic actuation principles for the knee joint
Here n is the number of links, j the number of joints and fk the number of DoFs
of joint k. The chain is statically determinate for m = 0 and indeterminate for
m < 0. The chain may move for m ≥ 1 where m denotes the number of independent
motion variables that governs the motion of the mechanism.
The first solution approach K1 shown in Figure 2.12a is the conventional design
which is well-known from nearly all humanoid robots. The electric motor (4) is
placed near the hip joint and connected to the gear unit (5) in the knee axis via
the synchronous belt drive (6). Similar configurations with a coaxial assembly
of motor and gear in the joint axis are not considered because of unacceptable
dynamic properties. Placing the motor near the hip joint significantly reduces the
thigh inertia, yet the gear unit remains coaxially with the articulated joint axis.
Thus, this solution is unfavorable because the high joint loading requires a large
and rather heavy gear unit. Compared to a motor-gear unit situated in the knee
joint, thigh inertia is reduced only marginally, and a considerable portion of the
hip joint performance must spent on accelerating the heavier thigh.
Solution idea K2, Figure 2.12b, complies with a claim in [51]. The structure of
the mechanism9 is RSUR with mobility m = 1. It comprises a rotary motor-gear
unit (7) located next to the hip joint. The knee (3) is driven by a parallel-crank
9 The common notation for lower kinematic pairs is as follows: R – revolute pair (f = 1); P –
prismatic pair (f = 1); H – helical pair or screw pair (f = 1); U – universal pair (f = 2); C –
cylindrical pair (f = 2); S – spherical pair (f = 3). If the joint is actuated, the corresponding
letter is underlined, e. g., R or H.
(a) Relative torque capacity (b) Overall reduction ratio

100 80
max(N ) = 71.6
90 70
Relative torque capacity [%]
80
60
Reduction ratio N [–]

70 Peak torque
Continuous torque
60 50
Gait pattern
Human (0 °/s)
50 40
Human (210 °/s)
40 30
30
20
20
10 10
0 0
-10 10 30 50 70 90 110 -10 10 30 50 70 90 110
Knee angle ϕknee [deg] Knee angle ϕknee [deg]
Figure 2.13: Torque-speed characteristics of the knee joint mechanism. The knee joint
angle definition complies with Figure 2.12a. The human torque capacity data in plot
(a) are adapted from [142].
mechanism. The coupler link (12) comprises spherical bearings (9) at its ends10 .
Kinematic singularities limit the angular range of the gear output to less than half
a rotation. Furthermore, the four-bar mechanism induces high bending moments
in the crank and follower links and high tension/compression forces in the coupler
link. The output bearing of the motor-gear unit must therefore show high load
carrying capability and tilting rigidity.
Solution approach K3 illustrated in Figure 2.12c is a slider-crank mechanism.
The mechanism has the structure PSUR with mobility m = 1 and is similar to the
design proposed in [35]. The linear guide (10) is fixed to the thigh (1) and connects
to the shank (3) by the coupler link (12) which employs spherical bearings (9) at
its ends. The actuator of the linear carriage (7) is either based on a ballscrew or
planetary roller screw, and omitted here for clarity.
Solution approach K4, Figure 2.12d, proposes another four-bar mechanism.
Assuming the thigh (1) as the fixed member of the linkage, the linear actuator (11)
comprises the two intermediary links of the mechanism. The linear actuator (11) is
universally mounted (8) in the thigh (1) and shank (2) and acts as length-variable
coupler. Hence the structure is SHUR with mobility m = 1. As with the two
previous approaches, this solution shifts the large proportion of the drive’s mass
close to the hip and knee joint performance is increased while the thigh moment
of inertia is kept low. Therefore, this solution approach is adopted in the design of
the robot. The torque-speed characteristic of this “muscle-like” mechanism K4 is
nonlinear and corresponds to the human knee [142]. As illustrated in Figure 2.13,
the torque depends on the link position and has its maximum at around 55°, which
is advantageous for typical gait patterns of the robot. Conversely, the maximum
speed increases at a stretched leg configuration, where they are needed.
10 Spherical bearings simplify the mechanical design but allow rotation of the coupler link
about its axis. Since this rotation does not affect general mobility of the mechanism, it is
considered as a passive or redundant degree of freedom and neglected.
2.6.3 Ankle Joint Actuation

The two-DoF ankle joint connects the foot and shank of the robot. Its motion
must allow for shock absorption and weight support and ensure reliable foot-
ground contact during stance. The dynamic performance of the ankle joint not
only affects gait velocity, rather, it must allow the precise control of the foot
torques which are important control variables for robot stabilization [18]. Hence
the drive mechanism must be free from stick-slip, it should show only minimal
backlash and have good friction behavior (preferably rolling friction). Finally, an
appropriate range of motion of the ankle joint is necessary as the dexterity of the
foot determines step length and the ability of the robot to step over obstacles and
to climb stairs.
Related Work
The ankle joints of almost all full-size humanoids are actuated in a serial ar-
rangement through Harmonic Drive gears, forming a universal joint. plantarflex-
ion/dorsiflexion is followed by the adduction/abduction axis which are fixed to
the shank and foot, respectively. This structure allows to easily reduce shank
inertia: as granted in a patent by Honda [50], the plantarflexion/dorsiflexion
motor can be placed near the knee joint without any additional transmission but a
synchronous belt. Most other robots have the plantarflexion/dorsiflexion motor
situated close to the ankle joint. The adduction/abduction motor is typically
mounted concentrically in the joint axis [50, 94], on the ankle joint segment [127,
130, 140] or — less frequent—near the knee joint [119]. In this case an angular
gear is necessary to redirect the motor axis. Another patent application by Honda
[118] describes a low-inertia serial structure, where both ankle joint motors are
mounted on the thigh segment. The plantarflexion/dorsiflexion gear unit is lo-
cated on the shank below the knee joint and connects to the ankle joint through a
parallel-crank mechanism. The adduction/abduction gear unit is concentric to the
joint axis and driven by the motor through synchronous belt transmissions and
an angular gear. Ogura et al. [130] propose a serial three-DoF ankle joint for the
robot Wabian-2 LL with internal/external rotation axis. The kinematic redundancy
can be utilized for avoiding obstacles or passing through confined spaces.
Only few robots have a parallel structure implemented in the ankle joints.
Espiau and Sardain [35] develop a two-DoF parallel slider-crank mechanism
with linear actuators (structure11 2-DoF 2 − PSU) for the robot BIP 2000. The
linear drives employ planetary roller screws. They are fixed to the shank and
connect to the feet with coupler links. Linear bearings keep the screws free from
radial loads. A patent application filed by Honda [51] claims a parallel mechanism
based on rotary drives. Two Harmonic Drive gears are mounted coaxially in the
knee joint axis. The output shafts on the left and right hand side of the knee
drive the ankle joint through spatial four-bar mechanisms (2-DoF 2 − RSU). The
11 Parallel mechanisms with identical kinematic chains are denoted as n − hchaini where n is the
number of kinematic chains, and hchaini denotes the type of the kinematic chains, with the
last letter denoting the kinematic pair at the end-effector (cf. Footnote 9 on page 54). Since
the information on the type of kinematic chains and the number of actuators is not sufficient
for determining the mobility of the mechanism, the latter is stated explicitly.
(a) Ankle joint speed (b) Ankle joint torque

6 200
5
4 150
3
100
Speed ωl [rad/s]
Torque τl [Nm]
2
1
50
0
-1
0
-2
-3 -50
Adduction/abduction Adduction/abduction
-4 Dorsiflexion/plantarflexion Dorsiflexion/plantarflexion
-5 -100
11 12 13 14 15 16 17 11 12 13 14 15 16 17
Time t [s] Time t [s]
Figure 2.14: Torque-speed requirements of the two-DoF ankle joint at a walking speed
of 5.55 km/h
motors are mounted on the thigh close to the hip joint and connected to the
gear input shafts by synchronous belt transmissions. Regarding leg dynamics
and lightweight design, this is undoubtedly an excellent solution. Gienger [46]
develops a parallel mechanism by for the robot Johnnie which has the structure
2-DoF 2 − HSS. Ballscrew-based linear drives are mounted on the shank and
drive two push rods which connect to the foot via coupler links. Each push rod
is guided by linear ball bearings to keep the ballscrew free from radial loads.
Due to the parallel structure the joint torques are distributed almost equally
among the two motors, which is the main advantage over Harmonic Drive-based
serial structures. Furthermore, torque control of the ankle joint which is an
important mechanism in balance control of the robot,benefits from the very low
rolling friction without stick-slip and the backdrivability of the ballscrews. During
operation the mechanism showed a few shortcomings. The angular workspace is
designed for normal walking and “jogging.” But self-collisions limit the maximum
step length and the robot’s ability to pass over obstacles and climb stairs: the push
rods move in the longitudinal axis of the shank and collide with the robot structure
at certain postures of the ankle joint. The push rod bearing arrangement has a
cantilever structure where the free length of the push rod depends on the ankle
joint posture. Especially for larger plantarflexion/dorsiflexion angles the free
length of the push rods increases and increases the inclination of the coupler links
with respect to the push rods. The induced bending moments cause considerable
deformations of the push rods and severely limit the lifetime of the push rods and
their bearings.
Solution Approaches
Figure 2.14 shows the torque and speed characteristics of the ankle joint at an
approximate walking speed of 5.55 km/h, obtained from the dynamics simulation
model (see Section 3.3.1). Because of apparently different lateral and sagittal
dynamics of the robot, both axes differ in their torque-speed characteristics. The
peak torques in the adduction/abduction and plantarflexion/dorsiflexion axes are
75 Nm and 175 Nm, respectively, the maximum joint velocities are approximately
1 rad/s and 5.5 rad/s, respectively.
If the ankle joint is actuated in a serial arrangement, a rather powerful drive
is required for the plantarflexion/dorsiflexion axis while a less powerful one is
sufficient to drive the adduction/abduction axis. Although it is possible (and
reasonable) to design tailored drives for both axes which differ in the motor per-
formance and gear output, it is unlikely to achieve significant weight reductions.
During stance the kinematic configuration of the leg in ground contact changes
and the ankle joint becomes its root joint. The gear units in both axes of the
ankle joint must show high torsional stiffness in order to avoid oscillations of the
stance leg and to minimize deflection (cf. Footnote 6). Hence the gear size in the
adduction/abduction axis cannot be reduced accordingly.
On the other hand, parallel actuation concepts can reduce the motor peak
torques significantly. Since the peak loads of both axes do not occur simultaneously,
the loads are distributed almost equally among the actuators. Thus, the motors
can be identical and smaller in size. Due to positive experience with ballscrew-
based parallel mechanisms in the ankle joints of the robot Johnnie, all solution
approaches considered in the following employ actuators based on ballscrews or
planetary roller screws. Figure 2.15 presents three possible kinematic structures
for the ankle joint that are based on linear actuators.
Solution approach A1 illustrated in Figure 2.15a allows to place almost the
entire drivetrain on the thigh. The linear carriages (10) are situated on either sides
of the thigh and driven by actuators based on ballscrews or planetary roller screws.
For clarity the screws are omitted in the figure. The carriage (10) drives the foot
(3) via the coupler links (11) and (12) and the ternary link (13) which rotates about
the knee joint axis (4). According to Eq. (2.5), the mechanism has two DoFs and the
structure is 2-DoF 2−PSSRSS. This solution allows an extremely low-inertia design
of the legs since most transmission components (in particular, the motors, screws
and linear carriages) are housed on the thigh (1). Only the coupler links (12) are
assigned to the shank (2). However, the large angular workspace of the knee joint
impairs the feasibility of this concept. A rather large stroke of the linear guides (10)
is needed in addition to the ankle joint motion to compensate the knee motion.
Moreover, the ternary link (13) may cause singularities at larger knee angles.
Solution approach A2,Figure 2.15b, has originally been proposed in [113]. Its
kinematics is similar to the knee joint mechanism presented in Section 2.6.2. The
linear actuators (9) act as length-variable couplers and are connected to the shank
(2) and foot (3) by the universal joints (7), resulting in the structure 2-DoF 2−UHU.
The mechanism is designed such that only tension and compression loads act upon
the drives, enabling the lightweight design of all components. The structure of
the drivetrain is rather simple and built from only a small number of components.
The linear actuators (9) are identical in design to reduce manufacturing effort and
part diversity. The shank length limits the length of the linear actuators (9) and
thus the angular workspace of the mechanism. The entire drivetrain is housed
on the shank which detracts from the low-inertia leg design. On the other hand,
relocating the motors to the thigh is unlikely because the drivetrain would have
to span the gimbal-mounting of the actuator (9). Finally, the linear actuators (9)
11
10 1 1
1 4
4
4
8 7
6 2 10
13 8
9
2 12
12
8 2
7
5 7
5
5
3 3 3
(a) Solution A1 (b) Solution A2 (c) Solution A3
1 Thigh 5 Ankle joint (b = 2) 9 Linear actuator (b = 2)

2 Shank 6 Revolute joint (b = 1) 10 Linear carriage (b = 1)
3 Foot 7 Universal joint (b = 2) 11,12 Coupler link
4 Knee joint (b = 1) 8 Spherical joint (b = 3) 13 Coupler plane (ternary link)
Figure 2.15: Kinematic actuation principles for the ankle joint (knee joint actuator
omitted for clarity)
are exposed on the back of shank (2) and are easily be damaged in the case of
collisions. Solution approach A3 shown in Figure 2.15c was first proposed in [111]
and is similar to the design adopted to the robot Johnnie. The linear carriage (10) is
mounted on the shank (2) and actuated by a screw (not shown). It drives the foot
by two coupler links (11) which are connected to the carriage (10) and the foot
(3) by spherical joints (8) and universal joints (7), respectively. This configuration
results in the structure 2-DoF 2 − PSU. The main advantage of this solution is that
the screws are housed on the shank and protected from collisions. The workspace
can be larger than in the previous solution because collisions with the shank are
reduced by the slender design of the coupler links (11). Therefore, mechanism A3
is chosen for the ankle joints of the robot.
Until now, Figure 2.15 has considered only the principal structure of the mech-
anism. The motors and auxiliary transmissions to connect the motors to the
mechanism have been neglected and is worked out in the following. Figure 2.16
presents two variants of mechanism A3. Mechanism A3.1 shown in Figure 2.16a
is the simplest solution. The motors (6) are housed on the shank and connected
to the planetary roller screws (9) by synchronous drive belts (7). The posture-
dependent inclination of the coupler link (13) induces radial and moment loads
to the screw (9). To avoid non-axial loading of the screw, the linear bearing (10) is
used to accommodate these loads. The entire drive mechanism A31 is mounted
6
7
1
4
Oxyz 8
1
4
7
lPRS2
lPRS1
6
9
10
A1 A2
11
11 10
13
2 13 9
5 2 s 5
12 B1 K B2
3 12 3
f
(a) Drive mechanism A3.1 (b) Drive mechanism A3.2
1 Thigh 6 Motor 10 Linear carriage

2 Shank 7 Synchronous belt drive 11 Spherical joint
3 Foot 8 Angular gear 12 Universal joint
4 Knee joint 9 Planetary roller screw 13 Coupler link
5 Ankle joint
Figure 2.16: Possible drive mechanisms for ankle joint mechanism A3. For clarity the
mechanisms are projected into the drawing plane.
on the shank which corresponds to the design adopted to the robot Johnnie [47].
Although the motors are situated close to the knee joint to minimize shank inertia,
the effect on the resultant leg inertia is barely visible. Moreover, because of limited
installation space on the shank small-diameter motors with larger stack length
must be used to produce the required torques. These motors are normally heavier
than larger-diameter motors with shorter stack length at the same power rating.
Mechanism A3.2, Figure 2.16b, goes one step further. The motors are located on
the thigh (1), as close as possible to the hip joint. Power transmission from the mo-
tors (6) to the planetary roller screws (9) is accomplished by two synchronous belt
drives (7) which connect to angular gears (8). The angular gears (8) are coaxially
with the knee joint (4) and redirect the axis of rotation from the knee joint axis to
the longitudinal direction of the shank (2). The output of the gearboxes (4) finally
drives the screws (9). This design allows to place the heavier transmission com-
ponents close to the hip joint, reducing their inertial contribution to a minimum.
The larger available space in the hip-thigh area allows to use motors with larger
diameter and reduced stack length which yields additional weight savings at equal
or increased motor performance. The additional transmission components allow
for the supplementary adjustment of the reduction ratio, for example by changing
the configuration of the belt drive (7) or the angular gear (8). On the other hand,
the additional transmission components reduce the overall efficiency. In addition,
the angular gear (8) must be designed very carefully to avoid greater backlash on
the motor-side. Complexity and part count are considerably higher as opposed
to mechanism A3.1. The longer transmission distance between motors (6) and
screws (9) is expected to increase the overall mass of the mechanism and goes
along with a higher motor-side inertia Jm , which reflects through the gears to the
load-side by the square of the reduction ratio. In the final analysis, drive mecha-
nism A3.2 is selected to drive the ankle joint. The advantages of this solution are
particularly evident in the effects on global system dynamics due to the minimal
inertial contribution.
Using the plantarflexion/dorsiflexion and adduction abduction angles, ϕfl and

ϕadd , respectively, the relative Jacobian of rotation for the roller screws is obtained
(cf. Figure 2.16b). The strokes of the planetary roller screws to adjust a certain
posture (ϕfl , ϕadd ) are given by:
!
kr OK + Asf (ϕfl , ϕadd )r KB1 − r OA1 k
l PRS = . (2.6)
kr OK + Asf (ϕfl , ϕadd )r KB2 − r OA2 k
Here r OK denotes the vector from the origin of the shank coordinate system, Oxyz ,
to the ankle joint pivot point, K. Asf is the transformation matrix from the shank
into the foot frame, r KB,i are the vectors from the ankle joint to the foot hinging
points (i = 1, 2) and r OA,i the vectors from the shank origin to the hinging points of
the linear carriages. The corresponding screw rotations for given strokes l PRS are
2π
ϕ PRS = l , (2.7)
Ph PRS
where Ph is the roller screw pitch. The relative Jacobian that relates the angular
speed of the screws, ϕ̇ PRS , to the joint angles, ϕfl and ϕadd , is calculated from
Eq. (2.6) in a straightforward manner:
ϕ̇ PRS ϕ̇ PRS 2π l˙PRS

= = . (2.8)
q̇ (ϕfl , ϕadd )T Ph (ϕfl , ϕadd )T
Finally, the reduction ratio yields from Eq. (2.8), and the reduction ratios of
synchronous belt drive and angular gear, Nbelt and Nbevel , respectively:
ϕ̇ PRS
N = Nbelt · Nbevel · . (2.9)
q̇
The resulting speed-torque curves of the ankle joint motors are compared in
Figure 2.17 to the serial actuation of the ankle joint axes using conventional gear
units. Note that, since drive mechanism A32 spans the knee joint, the motor speeds
ω m are also a function of the knee angle, ϕknee . Without loss of validity, the knee
angle can be assumed constant during the mechanism design, hence the relative
Jacobian, ω m/ q̇ , is given by Eq. (2.8). For better comparison, the reduction ratios
for serial actuation match the mean ratio of the parallel mechanism (Navg ≈ 58).
The motors in the serial configuration show apparently different torque-speed
(a) Parallel mechanism A3.2, Navg ≈ 58 (b) Serial actuation, N = 58

3500 3500
Motor # 1 Adduction/abduction
Motor # 2 Plantarflexion/dorsiflexion
3000 b Operating point
3000 b Operating point
Motor speed |ωm | [1/min]

2500 2500
2000 2000
1500 1500
1000 1000
b
500 500
b
0 0
b b
0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
Motor torque |τm | [Nm] Motor torque |τm | [Nm]
Figure 2.17: Torque-speed diagrams for the ankle joint motors in serial and parallel
configuration, cf. [113]
requirements. While the plantarflexion/dorsiflexion axis is subjected to high

torques over a wide speed range, the adduction/abduction motor has to deliver
rather low torques at considerably lower speeds. Thus, the motors differ clearly in
their output power. At a first glance it seems that the motor power increases for
the parallel configuration. The speed ranges of both motors are equal while the
motor torques remain comparable to the serial setup. However, the continuous
operating points, determined by the average speed and torque (cf. Figure 3.2)
show a significant reduction of the average motor loads.
2.6.4 Toe Joint Actuation

The toe segment is the end link of the leg. The consistent lightweight design of
the toe joint is crucial to keep the resultant leg inertia low. The related work on
actuated and passive toe joints is presented in Section 2.2.6.
Figure 2.18 shows four possible actuation principles of the toe joint. Solution
idea T1 is the standard design using a Harmonic Drive-based rotary actuator
mounted coaxially with the toe joint axis, Figure 2.18a. This design is adopted
to the robots H6 and H7 [128] and, presumably, Toyota’s running robot [178]. In
order to meet the weight and space restrictions the performance of the rotary
actuator must be minimized and exceeding reserve capacities must be avoided.
Although simple in design and kinematics, the anthropometric proportions (i. e.,
the joint position) cannot be fully kept because of the dimensions of the rotary
actuator. Solution idea T2 shown in Figure 2.18b employs a rotary actuator (5)
that is situated close to the knee joint on the shank segment (1) to keep its inertial
contribution low. The output torque is transmitted to the toe joint (4) by the
coupler link (9), forming a four-bar mechanism. Solution idea T3, Figure 2.18c,
uses a linear actuator (6) on the shank (1) to drive the toe joint (4) via a coupler link
(9). The kinematics are similar to the mechanism proposed by Yoon et al. [208].
Both designs allow to locate all major components of the drive mechanism close to
2.7 Sensor Layout 63
6 7
5
1 8 1 1
8
1 9
3 3 3 9 3
8 7 4
5 8
2 2 4 2 4 2
(a) Solution T1 (b) Solution T2 (c) Solution T3 (d) Solution T4
1 Shank 4 Toe joint (b = 1) 7 Hydraulic cylinder

2 Foot 5 Rotary actuator (b = 1) 8 Spherical joint (b = 3)
3 Ankle joint (b = 2) 6 Linear carriage (b = 1) 9 Coupler link
Figure 2.18: Kinematic actuation principles for the toe joint
the knee joint, reducing leg inertia. Hence the anthropometric joint position is fully
kept. However, the actuating force must be transmitted over the two-DoF ankle
joint, so that the toe joint actuator must be capable of compensating the ankle joint
motion. However, the large angular workspace of the plantarflexion/dorsiflexion
axis complicates the realization of solutions T2 and T3. Finally, Figure 2.18d
illustrates solution T4 which overcomes the difficulties associated with the ankle
joint motion. Here, the toe joint is driven by an electro-hydraulic actuator with two
linear cylinders in a master-slave configuration. The driving cylinder is located
on the shank. It is actuated by an electric motor through a screw drive. The
driven cylinder is installed between the forefoot and toe (hallux). Connecting
both cylinders by a flexible hose allows to theoretically install the driving cylinder
on the pelvis or the upper body, providing the most effective reduction of leg
inertia. Despite high power density at compact dimensions, hydraulic systems can
develop leakage and the general difficulties with hydraulic systems mentioned in
Section 2.5.1 apply, too. The toe joint is therefore designed with a standard rotary
actuator, according to solution idea T1. Although other mechanisms may have a
smaller impact on leg inertia, the chosen design is advantageous as its kinematics
are most simple and there is no interference with other joints.
2.7 Sensor Layout

The sensor system must support the implementation of model-based control
algorithms. The variables to be measured are determined from the equations
of motion of a rigid multibody system. Further, it is necessary to disambiguate
the contact states of the heel and toe segments. Finally, additional variables are
required for motor control. Hence the following variables are to be measured for
calculating the manipulated variables and controlling the servomechanisms:
1. Vector of generalized coordinates, composed of:
a) Position and orientation of a frame of reference that is rigidly connected
to any of the robot’s bodies with respect to an inertial frame. The upper
body, which is the root link for the kinematic chain in the simulation
and control of the robot, is chosen.
b) Joint angles with respect to the previous link
c) Joint angular velocities with respect to the previous link
2. External forces, i. e., floor reaction forces and moments
3. Contact state of the feet
4. Additional variables required for field-oriented motor control:
a) At least two motor phase currents
b) Angular position of the permanent magnet rotor with respect to the
stator windings
The requirements on the inertial measurement system used to estimate the up-
per body position and orientation are discussed in Section 2.7.1. Sections 2.7.2
and 2.7.3 define the requirements on the joint sensors and the force/torque sensors,
respectively.
2.7.1 Upper Body Position and Orientation

The orientation and angular velocities of the upper body are essential control
variables for keeping the robot in balance [18]. They are estimated by an iner-
tial measurement system (IMS) which is one of the most important sensors: its
accuracy and signal quality considerably affect the performance of the balance
controller since the upper body orientation enters into the calculation of the foot
positions. At worst, the foot-ground contact occurs too early because of inaccurate
and noisy measurements which can easily destabilize the robot.
The IMS works in a closed control loop. The overall bandwidth of the sensor
must be 20 to 30 times the bandwidth of the walking gait, so that the closed loop
system is not impaired by the sensor dynamics. Assuming a step frequency of
approximately 2 Hz, hence the bandwidth of the IMS must be at least 40 Hz to
60 Hz. The dynamics of the robot can be separated into fast and slow subsystems.
The fast dynamic corresponds to the motors and the torque-controlled ankle joint.
The slow dynamic is due to global system dynamics, represented by the upper
body DoFs. Ideally, the IMS should have a data output rate that corresponds to
the sampling rate of the balance controller (approximately 1 kHz) but lower rates
are acceptable because of the slow upper body dynamic.
Most biped robots employ inertial measurement system that are based on
MEMS12 accelerometers and MEMS gyroscopes, for example [53, 95, 110, 139,
153, 178]. MEMS attitude sensors are widely used in automotive and consumer
applications, however, the performance of these sensors is not always sufficient to
meet the specifications required in many advanced applications, such as robotics.
Notwithstanding, due to the high investment and comparably bulky measure-
ments only few robots employ IMS based on fiber-optic gyroscopes. For example,
Honda’s robots P2 and P3 are equipped with laser gyroscopes [65]. Takahashi
et al. [179] develop an inertial measurement system composed of three fiber-optic
gyroscopes and three servo accelerometers for a modified version of HRP-2 at the
Digital Human Research Center of AIST. For improved attitude estimation, the
12 Micro-electro-mechanical system
authors propose to take the planned translational acceleration of the robot into
account for estimation of the gravity vector. Gienger et al. [48] develop an IMS
for the robot Johnnie which is composed of three MEMS gyroscopes (measuring
range ±150 deg/s) and three MEMS accelerometers (measuring range ±2g). The
orientation of the upper body is estimated using complementary filtering of the
gyroscope and accelerometer signals. The IMS transmits the analog sensor signals
to the central control unit where the upper body orientation is calculated after
digitization. However, the transmission of the analog signals over a longer distance
results in rather noisy signals.
In order to define the requirements on the inertial measurement system, the
effects of various measurement errors on the walking stability are analyzed using
the robot simulation tool (cf. Section 3.3.1):
1. Sensor bandwidth
2. Noisy measurements
3. Data update rate
4. Sensor alignment
First, the overall sensor bandwidth is limited to 80 Hz13 and 20 Hz. The data
output rate of the IMS conforms to the sampling rate of the balance controller
of 1 kHz. The robot walks stably in both simulations without visible deviation
from the intended path. In the next study analyze the necessary quality of the
sensors is analyzed by adding white noise to the accelerometer and gyroscope
signals before sensor fusion. Using these noisy signals, the robot walks stably
but drifts laterally from its original path. If the noise level is reduced by 5 times
oversampling, the lateral drift of the robot is reduced significantly and barely
visible. Note that in the case of oversampling the individual sensors are sampled
at 1 kHz but the data output rate of the IMS goes down to 200 Hz, hence the upper
body orientation and angular velocities are updated every fifth controller cycle.
Finally, the influence of alignment errors is analyzed by adding constant angular
offsets about the x-axis so that the upper body seems to be tilted sideward. The
gait is stable at an alignment error of 0.2° but the robot deviates laterally from the
intended path. The stability problems get more emphasized with larger alignment
errors. The ground contact occurs unexpectedly at an angular offset of 2.0° which
results in a jerky gait and significant lateral drift due to unexpected impacts. After
a few steps the robot tips over and falls down. The same behavior is observed
during actual experiments with the robot Johnnie and leads to the conclusion that
the balance controller is very sensitive to alignment errors of the IMS. In summary,
the following requirements on the IMS are derived from these results:
– Increase in bandwidth of single sensors is not necessary
– High measurement accuracy
– High signal quality (minimal noise)
– Low offset (long-term bias) is achieved by heading aiding, e. g., by using
optical flow
– Reliable adjustment of mounting misalignment, either in hardware or soft-
ware. Reliable alignment procedure at power-up
13 This configuration corresponds to the single bandwidth of the MEMS gyroscopes used in the
IMS of Johnnie.
– Due to slow global system dynamics, the data update rates of the IMS can be
many times higher than the sampling rate of the balance controller
2.7.2 Joint Sensors

The joint sensors acquire the actual position and speed values of the joints. In
addition, the motor controller must have an absolute position value of the per-
manent magnet rotor with respect to the stator windings and at least two motor
phase currents available for electrical commutation [166]. The rotary position
sensors are important for the control quality since their resolution and accuracy
affects the positioning accuracy, speed stability and bandwidth of the servo loop.
Instability of the speed control loop can be due to mechanical influence from the
drivetrain (e. g., distinct or unmodeled compliance) or low-resolution signals of
the position feedback device. Conversely, higher resolution and accuracy reduce
disturbances in the motor current, hence the motor operates more quietly with
less dissipation. Two architectures for closed-loop motion control systems exist.
They differ by the location of the feedback device in the drivetrain which directly
affects the quality of the feedback signals.
The indirect measurement of the joint angles and angular velocities relies on
an incremental rotary encoder on the motor shaft. The simplicity in system
integration and signal processing is supposed to be the main reason why this
technique is most prevalent to humanoid robots, examples include [47, 95, 104,
127, 131]. With motor-side position feedback the angular position of the motor
shaft is measured and the joint position is inferred through knowledge of the drive
mechanism between the encoder and the articulated joint. A repeatable reference
point signal within the joint workspace is utilized to obtain the absolute position.
The joint velocity is obtained for instance by numerical differentiation of the
position signal. The absolute angular position of the motor shaft for commutation
is obtained in the same manner and requires a magnetic alignment procedure
at power-up. Due to the incremental measuring principle, the zero points of
the joints must be determined at every power-up which makes it difficult to
start the robot from arbitrary postures. The angular resolution of motor-side
feedback depends on the gear ratio which is typically high. On the other hand,
the measuring accuracy is affected by mechanical factors such as backlash and
transmission errors of the drive mechanism. Furthermore, elastic deformations of
the gear and joint under load cannot be detected which severely limits the range
of stable loop gains, along with the overall servo performance [197].
Few robots adopt the direct measurement of the joint positions and speeds, for
example [7, 188]. Here the actual position and speed values of a joint are obtained
by load-side feedback using an absolute angular measurement system. Measuring
the joint variables on the output side of the servomechanism avoids drivetrain-
induced errors like backlash, hysteresis and other transmission inaccuracies. Also,
nonlinearities and compliance in the drivetrain can be compensated within the
low-level motor control loops. Using absolute feedback devices, the joint positions
must be calibrated once as described in Section 3.11.1, eliminating homing or
calibration at power-up. Thus, the robot can (theoretically) start from arbitrary
postures, which is vital to real-life scenarios. However, the effect of compliance on

the closed-loop performance must as well be considered for load-side feedback.
As the controller will encompass unmodeled torsional resonance modes [197],
substantial oscillations can arise from different responses of the motor and load
side (see Footnote 6).
Unlike most full-size humanoids, each joint of the robot is equipped with an
absolute angular sensor for direct measurement of the joint angles. The afore-
mentioned advantages of load-side measurement are particularly evident for the
knee and ankle joints which are actuated by complex drive mechanisms with
nonlinear transfer behavior, see Sections 2.6.2 and 2.6.3, respectively. An addi-
tional incremental rotary encoder is implemented on the motor shaft to provide
high-resolution position signals for the electronic commutation of the PMSM. If
necessary, this encoder can also serve as a fallback sensor for motor-side position
feedback. To improve operational security and to prevent the robot from self-
destruction, each absolute angular sensor is complemented by a photoelectric fork
sensor to limit the allowed angular workspace. The angular range of each joint
is determined by a cam mounted on the output flange. In case of exceeding the
allowed workspace, the PWM14 signals from the motor controller to the inverter
are interrupted by the motor controller module with minimal software interaction.
2.7.3 External Forces

The orientation of the upper body is maintained by controlling the contact torques.
In addition, active compliance for shock attenuation at initial contact is imple-
mented by force control of the normal contact forces [18]. The precise measure-
ment of the ground reaction forces and moments is therefore essential to the
control system of the robot. For most of the implemented control approaches
(cf. Section 1.1.2) it is sufficient to determine the vertical force fz and the lateral
moments mx and my by measuring the vertical ground reaction forces at three
different points of the contact plane. More sophisticated control strategies, on the
other hand, further demand for the tangential forces, fx and fy , and the moment
about the vertical axis of the contact plane, mz . For example, these measurements
facilitate the compensation of leg tension during double support. Since the robot
should provide a high flexibility in the implementation of model-based control
strategies, the full contact wrench (forces fx , fy , fz ; moments mx , my , mz ) is mea-
sured by a six-axis force/torque sensor (F/T sensor) situated within the load path
between forefoot and ankle joint.
Most full-size humanoid robots incorporate six-axis F/T sensors in their feet.
Most of them use commercial sensors—either off-the-shelf or customized [65,
95, 104, 131] but a few robots employ custom developments. These sensors are
designed either as stand-alone devices [46, 99, 140] or as integral parts of the
foot structure [44]. The only full-size humanoid without F/T sensors in the feet is
Toyota’s running robot [178]. While most robot developers rely on well-known
design concepts for multi-component force sensors, Nishiwaki et al. [128] adopt
different sensor concepts to the robots H6 and H7. The robot H6 has a three-axis
14 Pulse-width modulation
5
4 1 Ankle joint flange
2 Foot connecting flange
1 3
3 Shear beam
2 4 Strain gauge
5 Membrane
0 10 20 [mm]
Figure 2.19: Force/torque sensor of the robot Johnnie (adapted from [46])
F/T sensor composed of six load cells. Four load cells are located in the toe and two
in the forefoot. A six-axis F/T sensor with parallel support structure is developed
for the robot H7 [125]. Using such designs, the force metrology can be tightly
integrated to the feet which results in a rather lightweight and compact structure.
However, such sensors often suffer from limited accuracy and repeatability due to
hysteresis effects caused by backlash and friction in the load cells.
The physics of biped walking yield high vertical ground reaction forces and
lateral moments about the ankle joint axes but comparatively small forces and
moments in other axes. In addition, high impact transients up to a multiple of
the robot’s weight occur when the foot hits the ground (cf. Section 2.3). A market
survey has shown that commercial six-axis F/T sensors with appropriate measure-
ment range are rather bulky and heavy because of the industrially-ruggedized
transducer body made of steel. Moreover, the measurement ranges are equal
in all axes which results in a rather low sensitivity of less loaded axes. Most of
these sensors require external signal conditioning electronics which are usually
too bulky and do not integrate with the electronics architecture. Although the
design, manufacturing and calibration of multi-component F/T sensors are very
elaborate, the characteristics of the ground reactions suggest the development of a
customized solution based on the Johnnie-design.
Before actually defining the architecture of the robot’s F/T sensors, the custom-
made six-axis F/T sensor developed by Gienger [46] is discussed. The monolithic
transducer body is made from a high-strength aluminum alloy and holds three
shear beams as sensing members. It is shown in Figure 2.19 for reference purposes.
Its measurement range is tailored to the ground reaction forces and moments
that have been obtained from simulations [109]. The sensor is very compact
(∅ 90 × 30 mm2 ) and lightweight, weighing approximately 290 g including signal
conditioning electronics. The monolithic transducer body avoids setting and
hysteresis effects. Thin membranes reduce cross talk by decoupling the beam
deflections to a far extent. Due to narrow tolerances, the transducer body was man-
ufactured by EDM15 wire cutting. Each shear beam holds four strain gauges which
are selected to match the elastic properties of the transducer material. Two op-
posing strain gauges are series connected, allowing the temperature-compensated
15 Electrical discharge machining
2 z
1 3 y
1 Membrane
2 Strain gauge
3 Shear beam
Figure 2.20: Schematic design of a Maltese cross-shaped force/torque sensor
acquisition of biaxial beam bending. The sensor is designed as a stand-alone com-

ponent with connecting flanges to the ankle joint and foot. Hence modifications of
the foot structure are possible without impairing the sensor’s function. Since the
strain gauges are very sensitive to surface leakage currents, the sensor is isolated
from the rest of the robot. However, the low compressive strength of the insulating
material caused mechanical failure of the insulating layer after a short operating
time. Signal conditioning and offset compensation is contained within the sensor.
The analog measuring signals are transferred to the central control unit on the
upper body, where the digitization and signal processing are carried out. Hence
signal quality and measuring accuracy are impaired by electromagnetic interfer-
ence. The sensor is calibrated with respect to a reference system located in the
middle of the foot support polygon. Although this approach is computationally
advantageous to the first implementation of the on-board computer system [46],
it complicates changes in the foot geometry.
In the following the fundamental design concept of the F/T sensor is developed,
including the measurement principle, the transducer geometry, the integration
into the foot structure, and the signal processing.
Because of extensive experience from previous sensor developments metal
foil strain gauges are chosen as sensing elements. They have small physical
size and low mass and can achieve a very high accuracy in static and dynamic
measurements. The resistance change is due to deformation of the surface to
which they are attached, which is measured by Wheatstone bridges either in half
or full-bridge arrangement. Other physical effects utilized for force measurement
are introduced in [187].
Figure 2.20 shows a principal sketch of a F/T sensor in the classic Maltese-cross
design. The transducer body holds four equally spaced and crosswise-arranged
shear beams as the sensing members. The shear beam geometry is optimized for
high sensitivity even of the minor loaded axes without detracting from the sensor
stiffness. Hysteresis and setting effects are avoided by the monolithic design of the
transducer body. To decouple beam deflection and maximize strain on the beam
surfaces, the shear beams connect to the outer sensor frame by thin membranes
acting as flexure hinges. For reduced manufacturing effort, the transducer body
is manufactured by milling process. Selective sensitivity of certain half bridges
for certain forces or moments is achieved by the four-beam design. Moreover, the
redundant setup with eight half-bridges can be utilized to improve measurement
accuracy. The application of forces beyond the yield point of the gauge pattern is
a typical failure mode of strain gauges. Overloading can easily occur during the
testing of new control strategies or sudden malfunctions. It may cause permanent
damage of the strain gauges and shear beams, finally yielding zero shift, non-
linearity and fatigue of the gauge pattern and transducer. Therefore, the F/T
sensor is equipped with an overload protection mechanism.
The F/T sensor is housed on the end link of the legs. A well thought-out
lightweight design is therefore essential and affects the measuring accuracy, as
well as the integration of the sensor into the foot structure. Significant weight
reduction is expected from the tight integration of the sensor into the foot. The
sensor frame acts an integral, load-bearing element within the foot structure,
eliminating the additional frame structure required for the integration of a self-
contained F/T sensor. Hence part count, bolted flange interfaces and, finally, the
weight of the entire foot are reduced. To this end, it is necessary to optimize the
load-bearing sections of the sensor body for flexural and torsional rigidity. To
ensure a high measuring accuracy the deformations of the sensing elements and
load-bearing sections must be mechanically decoupled. Further weight savings
can be achieved by the intelligent placement of the sensor within the foot structure.
In almost all humanoid robots the F/T sensor is situated immediately below the
ankle joint. This arbitrary placement results in different moment arms of the
forces acting on the heel and toe contact elements and yields a higher moment
loading of the sensor. Shifting the sensor towards the center of the foot contact
area, the moment loading is reduced while the resolution is increased. Hence the
same moment arms account for contact forces acting on the heel and toe, similar
to a claim in a patent application by Honda [183].
A/D conversion and signal conditioning are integrated to the sensor electronics,
allowing for higher cut-off frequencies of the anti-aliasing filters. In conjunction
with oversampling signal quality and bandwidth are improved. The ground reac-
tion forces and moments are then calculated by a microprocessor and transmitted
to the central control unit (cf. Section 2.8). The calibration should be independent
from the foot geometry to simplify future modifications of the feet. Hence the
sensor is calibrated with respect to reference system located at the center of the
transducer body.
In summary, the most important requirements on the F/T sensor are:
– Acquisition of the ground contact forces and moments from all three Carte-
sian coordinates
– Maltese-cross-shaped transducer body with monolithic structure to avoid
hysteresis and setting effects
– Shear beam geometry optimized for high sensitivity even of minor loaded
axes
– Load-bearing structure designed for tight integration within the forefoot
structure
– Sensor placement in the center of the foot contact area
– Integrated signal conditioning and A/D conversion
2.8 Computer Architecture 71
– Calibration with respect to reference system located at the center of the

transducer body
2.7.4 Contact State of the Feet

Due to division of the foot into forefoot and toe, the contact state of the foot
contact elements is ambiguous and cannot be determined reliably. However, for
stabilization of the robot it is very important to know the contact state of the feet.
It can be determined for example by measuring distance of the contact elements to
the ground. Kajita and Tani [86] apply ultrasonic distance sensors to the robot
MELTRAN for real-time detection of the ground profile. Gienger [46] proposes
tactile sensors based on potentiometers to determine distance to the ground.
Measuring the distance to ground provides more information and enables more
sophisticated control strategies. In a first step, however, it easier to integrate
discrete contact sensors to the viscoelastic elements of the heel and toe segments.
The contact state is thus measurable and allows, for instance, to implement event-
triggered switching of the gait patterns.
2.8 Computer Architecture

This section deals with the conceptual design and fundamental operational struc-
ture of the on-board computer system. The performance of the computer archi-
tecture is essential to the performance of the walking robot. It has to acquire all
sensor data, calculate the trajectories, stabilize the robot, and compute and output
the motor control variables. Having a total of 25 actuated joints and numerous
sensors for state feedback, the overall system is quite complex. Regarding the
modular design of the servo actuators, the implementation of a fully decentral-
ized computer architecture seems obvious. While such architectures have been
implemented successfully for many multi-legged walking machines [144], balance
control of biped robots requires a central controller instance. Because of highly
coupled kinematics and dynamics the control algorithms cannot be reasonably
divided into self-contained modules that run on distributed nodes.
In a first approach, a computer system with distributed nodes was implemented
to the robot Johnnie based on the hardware concept of two previous multi-legged
robots [144]. It was composed of a PC-based, external host system and six on-
board microcontrollers. All nodes are connected via six CAN bus lines. To
achieve a high sampling rate, trajectory planning and stabilizing control run on
the host PC, while low-level motor control and data acquisition are handled by
the microcontrollers. Using this computer system, a minimum cycle time of 4 ms
has been achieved, comprising 2.5 ms for calculation of the controller on the host
PC, 0.5 ms for data transmission via CAN, and 1 ms for sensor data acquisition
and processing on the microcontrollers. Experiments suggest, however, that the
sampling rate of 250 Hz is too slow for the chosen control concept [109], impeding
the implementation of faster gait patterns. In the refined version the robot is
fully controlled by an on-board PC. An external computer connects to the robot
via Ethernet to give basic operating commands and for monitoring purposes.
A customized PCI board interfaces the on-board PC to the sensors and motor
electronics. The board is situated in the upper body of the robot and connected to
the sensors and electronics with a rather complex and bulky cabling. Thus, the
cabling makes about 4.7 % of the overall mass [46]. Many other humanoid robots
employ centralized hardware architectures, for example HRP-2 [95], Wabian-2
[131] and H7 [127]. Most of them employ custom-developed interface boards for
hardware I/O16 . Nevertheless, an increasing number of biped robots employs
architectures with distributed nodes for low-level data processing and control. In
this case a key problem is the communication between the decentralized nodes
and the host PC. Most of these robots uses the CAN Bus communication protocol,
typically with multiple channels, for example HUBO [140], ARMAR-III [151] and
iCub [188]. Few robots employ other communication systems than CAN. The
prototype robot HRP-3P uses a custom-developed real-time Ethernet system [3,
92] but is migrated to CAN in the robots HRP-3 [96] and HRP-4C [93]. Sarker et al.
[161] propose a real-time control system for the robot MAHRU which is based on
IEEE 1394 (FireWire). DLR’s lightweight robots LWR-II and LWR-III and the two-
arm system Justin have a modular computer system with one controller module
per joint. SERCOS with fiber-optic cables is used as the communication system at a
rate of 1 kHz [68]. All electronics are fully integrated to the joint modules [67, 137].
The upper-body humanoid Robonaut developed by NASA employs FPGA-based
motor controllers and bi-directional host CPU communication using synchronous
RS-485 drivers [14].
2.8.1 Design Concept

In order to make the robot computationally autonomous, the computer hardware
for trajectory planning, gait stabilization and low-level control is integrated to
the robot. As stated above, the highly coupled kinematics and dynamics of
biped robots demand a central controller instance. From a technological point of
view, however, it is feasible and advantageous to unload the central control unit
(CCU) from standard tasks: using a decentralized architecture, low-level motor
control and sensor data processing run in parallel on local controllers. Similar to
hierarchical structures in biological systems, sensor data is processed locally and
only relevant information is forwarded to the CCU. Thus, CCU, local controllers,
sensors and servo actuators form an “intelligent sensor-actuator network”17 . CCU
and local controllers are connected by a fast real-time communication system. By
unloading the CCU from hardware-related tasks, the local controllers abstract
the robot hardware from higher-level control tasks and thus simplify future
16 Input/output
17 Unfortunately, the term “intelligence” lacks a precise definition. Throughout this thesis,
“intelligence” is used in accordance with the definition in the field of sensor technology: an
intelligent sensor senses and digitizes the input signal and takes some action on it, e. g., am-
plification, zero error compensation, linearization and scaling. The sensor is also able to
execute auxiliary logical functions and instructions and the processed data are communi-
cated digitally to a control unit. Consequently, an intelligent actuator receives desired values
and configuration commands from a control unit and executes them independently. Again,
only actual values and status information are communicated to the control unit.
2.8 Computer Architecture 73
extensions of the robot. Malfunction of the on-board computer system may cause
serious harm to the robot and its environment. The mechanical and electrical
robustness of the computer hardware is therefore essential. Furthermore, the
complex point-to-point cabling of sensors and actuators is replaced by bus systems
for communication and energy supply. Cable harnesses for connecting the servo
sensors, actuators and drives to the local controllers can be kept short and mostly
do not have to be routed across moving joints. Decentralized architectures are thus
expected to reduce the cabling effort and improve electromagnetic interference.
Further decentralization of higher-level control tasks is not considered, so
that trajectory planning and control of system dynamics run on the CCU. Nev-
ertheless, this architecture allows for the implementation of different levels of
decentralization, ranging from more centralized signal processing concepts to
fully decentralized motor control. In the first approach only the current loop is
closed in the drive and all other loops are implemented in the CCU. The second
approach closes all control loops, including motor position, velocity and current,
in the drive. In summary, the functions of the computer components are:
Central control unit (CCU): trajectory planning and stabilizing control; support
the implementation of control architectures with different levels of decen-
tralization.
Distributed controllers: low-level motor control, including current, velocity and
position control, depending on the control architecture; evaluation of the
joint sensors; output control signals to the motor power electronics; parame-
terization of the motor controllers, e. g., for gain scheduling; data acquisition
from the force/torque sensor and inertial measurement system.
Real-time communication system: deterministic connection of CCU and local
controllers with low latency; transmission of actual and desired values, and
status and diagnostic messages; parameterization of the drives; switching
between different operating modes of the motor controller.
2.8.2 Central Control Unit

Using a PC-based platform, the central control unit has the highest possible
computing power that can be installed on-board. By using only off-the-shelf
components, the CCU can easily be upgraded to the performance requirements
of the stabilizing control and trajectory generator. The CCU is built from an
industrially ruggedized PC board that is supplemented by an interface board for
the communication system.
2.8.3 Decentral Controllers

The decentral controllers must provide all functionality for sensor data acquisition
and motor control to reduce the volume for external electronics, as well as the
size and number of cable harnesses throughout the system. The proposed system
requires the development of custom hardware because there are no commercial
controller modules available that provide the combination of interfaces to various
sensors and the communication system at compact measurements and low weight.
Briefly, the mandatory requirements for the decentral controllers are:
– Interface to the motor shaft encoder, link position sensor, limit switch,
current and temperature sensors of each joint
– Suitable to control three PMSM at an intermediate circuit voltage 80 V at up
to 25 A (continuous) and 50 A (peak)
– Cycle times of the position control loop as low as 1 ms
– Interface to the force/torque and foot contact sensors
– Interface to the inertial measurement system
– Small outline to allow for positioning close to the servo actuators
– Modular design for reduced development effort
– Low power consumption
– Interface to the real-time communication system
The full integration of the motor controllers and inverters into the servo actuators
would consistently follow the idea of a modular hardware design as discussed in
Section 3.4. Also, the cabling of the servo actuators would be improved. The huge
development effort, however, is beyond the scope of this project. Therefore, it is
proposed to integrate the low-level motor control co-located with the actuators.
2.8.4 Communication System

The real-time communication system is the only data interface of the CCU to the
robot hardware. It must handle the entire data exchange between the CCU and
local controllers. The 25 motor controllers, two force/torque sensors and on IMS
are distributed equally among the local controllers. Then the communication
system must be able to handle data transfer between a total of nine local controllers
within the sampling time of the balance control loop of approximately 1 ms.
Multi-axis control applications like biped robots require the synchronicity of
all control loops for coordinated motion. Time delay in the control loops is
due to the computing time of the control algorithms and the duration of sensor
data acquisition, processing, and data transfer. High latency limits the sampling
time and can diminish control performance. The communication system should
therefore show minimal latency and small protocol overhead. Handling this large
data stream with minimal latency is even more difficult due to the small packaging
goals that limit overall electronics volume and geometry. For reduced weight and
cabling effort the communication system should show bus or ring topology. As
the local controllers require custom hardware, implementation effort should be
reduced by available driver software and hardware interfaces (ASICs or PCB
modules). The requirements on the real-time communication system connecting
CCU and local controllers include:
– Real-time capability and a high level of determinism
– Low cycle times
– Guaranteed, high bandwidth for data transfer to all devices within one
control cycle
– Synchronicity and minimum of protocol overhead to minimize latency in
the control loops
2.9 Chapter Summary 75
– Synchronization of the local controllers to the CCU time base

– Commercially available driver software and hardware interfaces for simpli-
fied system integration
– Small footprint of the hardware interface
2.9 Chapter Summary

This chapter presented the mechatronic design concept of the robot. The following
design guidelines for the development of the detailed design of the mechanical
and electronic subsystem were derived:
Mechanical configuration: redundant kinematic structure with 25 DoFs, includ-
ing actuated toe joints; anthropometric proportions of the robot as far as pos-
sible; foot dynamics concept with viscoelastic layers and hydraulic damper
for effective shock attenuation at initial contact; consistent lightweight de-
sign of the mechanical structure with high effective stiffness and natural
frequencies; low-inertia design of the leg apparatus for good dynamic per-
formance of the locomotor system.
Joint mechanisms: all joints are driven by servo actuators combining precision
gearing with a high-dynamic AC servo motor and sensors in a compact
housing; the hip joint is composed of three serial drives, forming a spherical
joint with all joint axes intersecting in a single point; for improved mass
distribution in the hip-thigh area, the knee joint is driven by a planetary
roller screw-based linear actuator acting as length-variable coupler; the
ankle joint is driven by a parallel mechanism composed of two spatial slider-
crank mechanisms, where the motors are mounted on the thigh, close to the
hip joint.
Sensor layout: supports the implementation of model-based control algorithms;
upper body orientation with respect to an inertial frame are measured by a
high-precision inertial measurement system based on fiber-optic gyroscopes
and MEMS accelerometers; joint angles and angular velocities are measured
by absolute angular sensors on the gear outputs; additional motor control
variables are measured by incremental encoders on the motor shafts; exter-
nal forces and moments are measured by customized six-axis force/torque
sensors in the feet; contact sensors in the heel and toe segments disambiguate
the contact state of the divided feet.
Computer architecture: the “intelligent” sensor-actuator network with central
control unit (CCU) improves system performance from a technological point
of view; the CCU accounts for trajectory planning and global system sta-
bilization; decentral controllers account for low-level motor control and
sensor data acquisition and signal processing; a high-bandwidth real-time
communication system connects CCU and decentral controllers in order to
minimize latency in the control loop.
3 Mechanical Design
Schnell und fein das kann nicht sein.
—My Mother
This chapter presents the design of the mechanical structures. Figure 3.1 shows a
photograph of the robot Lola, the key data are summarized in Table 3.1. Further
dimensions are given in Appendix B.
The development of a complex mechatronic system such as a humanoid walking
robot is an iterative and open-ended process of design and simulation. To develop
the mechanical system both structured and systematically, the design methodol-
ogy described in Section 2.1.1 is applied. According to Pahl et al. [138], the design
process divides into the four phases of task planning and clarification, conceptual
design, embodiment design and detail design. The two last phases, which are
subject of this chapter, are based on the requirements and fundamental design
concept worked out in the previous chapter. The process of mechanical design
and the tools used are introduced briefly in Section 3.1. Additional design rules to
achieve a consistent lightweight construction with good dynamic properties are
given in Section 3.2. The design calculations, Section 3.3, are based on the dynam-
ics simulation of the full system developed within the sub-project “Simulation
and Control.” They include the dimensioning of the motors and drive mecha-
nisms and the dimensioning of structural components which are addressed in
Sections 3.3.2 and 3.3.3, respectively. To reduce the complexity, the overall system
is divided into well-defined sub-assemblies which can be developed consecutively
and independently. These are, the
– joint actuators (Section 3.4),
– camera head (Section 3.5),
– trunk and pelvis (Section 3.6),
– arms (Section 3.7),
– legs (Section 3.8), and
– feet (Section 3.9).
Many of these sub-assemblies include the mechanical design of link segments,
some prominent examples are described in Section 3.10. After assembly the
kinematic and inertial parameters of the robot are identified. The associated
procedures are described in Section 3.11. Lastly, Section 3.12 concludes with a
discussion of the mechanical design.
3.1 Design Process

The design process comprises the development, evaluation and selection of several
design variants and the detailed geometric design of all mechanical components.
77
78 3 Mechanical Design
Figure 3.1: 25-DoF humanoid walking robot Lola

3.2 General Design Rules 79
Table 3.1: Hardware Specifications
General
Height 180 cm
Mass 60.88 kg
Total DoFs 25
Maximum walking speed (target) 5 km/h
Power supply external
Legs
7 DoFs (3 hip, 1 knee, 2 ankle, 1 toe)
Thigh length 440 mm
Shank length 430 mm
Foot height 105 mm
Distance between hip joints 246 mm
The 3D-CAD system CATIA V5 from Dassault Systèmes1 is adopted for most
of these steps. The robot components are modeled using the CATIA modules
Part Design (PDG), Generative Sheetmetal Design (SMD) and Generative Shape
Design (GSD). Stress analysis and strength assessment are mainly conducted with
the built-in structural analysis modules (GPS/GAS) which provide automatic
finite element meshing from native CAD data. CATIA’s digital mock-up (DMU)
module is applied to confirm the assembly of critical components and identify self-
collisions. In particular, the drivetrains of the knee and ankle joints are analyzed
thoroughly because the complex kinematics of these mechanisms with several
moving parts affect the entire surrounding construction.
The CAD model is organized in the style of the modeling strategies of the
guideline VDI 2209 [193] and resembles the tight coupling of mechanical design
and dynamics simulation. Each sub-assembly of the CAD model corresponds to
a body of the dynamics simulation model (cf. Section 3.3.1). The “root part” of
all sub-assemblies does not contain any actual part geometry but a local frame of
reference which is defined in accordance with the multibody simulation model.
The “root part” further comprises some characteristic elements (reference planes,
points, axes, etc.) that are primarily used to define assembly constraints between
components. This approach ensures consistency of the CAD and simulation
models and simplifies the acquisition of new inertia properties from the CAD
model.
3.2 General Design Rules

The fast-moving cantilever links of the legs yield high inertia forces and conse-
quently, high joint torques and bending moments in proximal link sections. Al-
though stresses are normally low, compliance is becoming more pronounced with
such link structures. In consequence, unwanted effects on system performance
(positioning accuracy, dynamic stability, static and dynamic fatigue strength) arise
1 http://www.3ds.com/
from inadequate link and joint stiffnesses. Therefore, the objective of lightweight
system construction demands for a methodical procedure, taking lightweight
design into account even at the conceptual stage. From the system’s perspective,
“conceptual lightweight construction” generates savings through functional inte-
gration using highly integrated components that combine several functions and
parts. Further, the use of lightweight materials such as high-strength aluminum
alloys (“lightweight material construction”) provides the basis for great weight
reductions, cf. Section 3.10. Moreover, the geometrical optimization of single
components (“design for lightweight construction,” cf. Section 3.3.3) facilitates
reduced cross sections with appropriate stiffness.
In the following, a number of general design rules are established in order to
achieve a sophisticated mechanical system with appropriate structural dynamic
characteristics:
Direct load transfer and distribution: direct load transfer to the main structure;
avoid redirection of forces; establish large-area application of loads; design
sophisticated load application and distribution structures for thin-walled
components; optimized support conditions of areas of load transfer; addi-
tional ribbing of flange joints; symmetric design of components; preferably
closed instead of open cross sections (“Monocoque-like design”); cambered
cross sections for larger load rating.
Preferably filigree structures: prefer thin-walled cross sections with stiffness
reinforcing ribs over solid bodies; optimize mass distribution of a compo-
nent to achieve the largest second moment of area (bending) and/or polar
moments of inertia (torsion) for a given mass of the component.
Avoid stress concentrations: avoid weak spots by allowing only smooth changes
in cross-sectional shapes.
Systematic stiffening in directions of maximum stress: develop a judicious sys-
tem of ribs or other reinforcements to achieve anisotropic stiffness wherever
appropriate.
Integral part design: reduce part count and assembly complexity; prefer milling
from the solid over attaching parts whenever possible; avoid overlap due to
bolted or riveted flange connections.
Flange connections: prefer rivets over screws if non-detachable connections are
acceptable; arrange flanged joints within the load path (preferably close to
the wall or within recesses); design for a direct force flux between the flange
members; prefer a larger number of uniformly arranged smaller screws
over a lower number of larger screws; prefer longitudinal over transverse
stressed bolt connections; use friction enhancement shims [34] for safe load
transmission even over smaller flange structures; combine bolted connec-
tions with positive locking for high joint stiffness and reduced screw count;
reduced offset between the bolt pattern and line of action of the applied
forces provides the most effective means of increasing the joint stiffness.
Damping enhancement reduces strain and stress amplitudes: different excita-
tion frequencies act on the robot structure which cannot be determined a
priori, depending on the gait pattern, control strategies and other parameters.
Thus, a very effective way to reduce structural vibrations is to systematically
3.3 Design Calculations 81
enhance damping, e. g., by applying viscoelastic elements [155].

Joint stiffness: link connections frequently contribute the dominant structural
deformations [154] and should therefore be designed as stiff, backlash-
free bearings; prefer adjusted bearing arrangements; prefer thin-section
bearings with larger diameter over smaller bearings with equal load rating;
use precision bushes with minimal clearance for plain bearings.
Although the goal is to tap the full potential of the design, there are some lim-
itations of this approach. Consistent lightweight design is only achieved if the
safety margins for unseizable boundary conditions are reduced. Hence the pre-
cise knowledge of all loads acting on the system (in magnitude and direction) is
required. In addition, the use of materials with guaranteed specifications and
preliminary tests to validate structural details are essential [101]. On the other
hand, the robot is planned as a multi-purpose research device. Depending on the
speed of locomotion, gait patterns and control strategy, the loads acting on the
system can differ significantly. Moreover, the robot must be able to move about
in unfamiliar space and/or over obstacles when operated with the autonomous
vision and path-planning system. To summarize, task uncertainty obviously goes
along with uncertainty of loading and must be taken into account in the design
of critical components. Therefore, the prototype investment castings of the link
segments (Section 3.10) and some other complex and cost-intensive parts are
designed for high-cycle fatigue to avoid failure.
3.3 Design Calculations

This section focuses on the sizing calculations that are carried out within the scope
of this work, which include:
1. Sizing of the actuators and their transmissions
2. Structural dimensioning of the links and joints
3. Elastodynamic analysis of the mechanical structure and drive mechanisms
Sections 3.3.2 and 3.3.3 address the dimensioning of the drivetrains and mechan-
ical structures on basis of static and dynamic loads. Static loads include forces
and moments under the most demanding operation conditions and correspond to
the maximum forces and moments the feet are able to transfer to the ground at a
certain robot posture. They represent the worst-case loading that the mechanical
system should be able to withstand without damage. Also, static loads are used
to determine the actual gear sizes and for worst-case finite element analyses. On
the other hand, dynamic loads that occur during normal operation are by far
more important and determined with a full dynamics simulation of the robot,
Section 3.3.1. These loads are used to determine the dynamic performance of
the actuators and gears. Moreover, they serve as constraint forces in the strength
analysis of mechanical components.
3.3.1 Calculation of Fundamental Design Data

The mechatronic modeling of the robot is a key tool for hardware and controller
design. While the static loads can be determined quite easily by analytical as-
sumptions, the dynamic performance data are obtained from this comprehensive
simulation model developed by Buschmann [19]. Considering the major effects
on bipedal locomotion, the robot is modeled as three coupled dynamical systems:
– Multibody dynamics (rigid body mechanics, drive friction, gear elasticity,
nonlinear kinematics in ankle and knee joints)
– Contact dynamics (unilateral, compliant foot-ground contact)
– Electrical motor dynamics
The stiffness and damping characteristics of the foot-ground contact are adjusted
by experiments with the robot Johnnie, so that the computed results can be ex-
pected to be close to reality [19]. Typical load profiles at a walking speed of
5.55 km/h are given in Appendix D.
The joint loads (torques and velocities), external forces (ground reactions) and
the constraint forces and moments of the links are very important data throughout
the design process. They are used for the dimensioning of the drive mechanisms
and structural components. Moreover, the angular ranges of typical gait patterns
are important constraints for the robot topology and link geometry. Conversely,
the topology of the kinematic chain, geometric dimensions of all links, motor
parameters, gear ratios, and the dynamic properties of the mechanical system
serve as input parameters for the robot simulation. After certain milestones of
the mechanical design have been completed, the inertia properties of the links
and actuators are calculated numerically from the CAD model and imported into
the robot simulation. After the balance controller has been adjusted to the new
parameters, new design data are calculated to be used in the dimensioning of the
mechanical system. In this way both simulation and CAD model are iteratively
refined and the final design of the robot is evolved. Finally, the mass parameters
of the real robot have been determined during assembly (see Section 3.11.2) to
further improve modeling accuracy. Hence the dynamics simulation can be used
as a reliable tool for developing and analyzing control system robustness and
performance before experimental verification on the real hardware.
3.3.2 Dimensioning of the Drive Mechanisms

Dimensioning the servo drive mechanisms and their components is a demanding
task due to constantly reversing motion and the highly variable loading of the
joints. The fast link accelerations at higher speeds of locomotion demand for good
dynamic performance. The actuators and drive mechanisms are selected, thus, for
– meeting the static and dynamic load requirements, while taking account
enough reserve capacity;
– sufficient drivetrain stiffness to avoid oscillations of the leg structure;
– enabling a good tracking and disturbance rejection behavior of the motor
controllers;
– good energy efficiency.
The iterative selection procedure for the drive mechanisms is illustrated in
Figure 3.2. This section focuses on the highlighted steps. The chosen procedure
allows the comparison of different motor-gear combinations with regard to power
density, dynamic performance, stiffness and power dissipation. Starting from
Pre-selection
Dynamics simulation
of overall system
Torque-based dimensioning
Elastodynamic analysis
Change configuration
Load-to-motor inertia ratio
Operating life calculation

of transmission components
no
Verification successful?
yes
End
Figure 3.2: Iterative selection procedure for the motors and drive mechanisms. This
section deals with the torque-based and elastodynamic dimensioning of the drive
mechanisms, and the effects of the load-to-motor inertia ratio on the performance of
the servo system.
motor-gear combinations that are expected to be capable of driving the joints,

the actual joint loadings are determined by the robot simulation. The static and
dynamic load requirements are then analyzed under the assumption of a rigid
drive mechanism. The torque-based dimensioning must balance the conflicting
requirements of minimal weight and overloading capacity: consistent lightweight
design implies that the servo actuators should not be oversized, at the same time
appropriate reserve capacity must be provided to cope with task uncertainty.
However, the finite torsional stiffness of the transmission components introduces
compliance in the drive mechanism between the motor and load. Such compliance
limits the bandwidth of the mechatronic servo system and can introduce resonant
oscillations. Hence the elastodynamic analysis of the drivetrain accounts for elimi-
nating or reducing resonances in the drive mechanism. Also, compliant drivetrain
components—in particular, the motor-load coupling—greatly affect the tracking
and disturbance rejection behavior of the servo system. To achieve good control
performance, the load-to-motor inertia ratio is considered as an indicator of pos-
sible stability problems. Finally, the operating life of transmission components is
calculated where necessary. If the load rating, operating life or dynamic behavior
of the drivetrain does not meet the requirements, the drivetrain configuration is
modified and the procedure repeats.
Torque-based Dimensioning
The fundamentals of motor sizing are well-known and beyond the scope of this
thesis, so that only a rough overview is given in the following. The reader is
referred to the relevant literature, e. g., [165]. Only the continuous and intermittent
load ratings of the servo motors are discussed with special regard to thermal
capabilities.
In order to dimension the servomechanism, the torques acting on the motor and
drive mechanism must be analyzed. Assuming a stiff connection of motor and
gear, the motor torque τm necessary to follow a given load profile is
τ τf ,l
τm = Jm N ω̇l + l + τf ,m sgn(ωm ) + sgn(ωl ) . (3.1)
N N
|{z} | {z }
=ω̇m =τf
In Eq. (3.1) the load-side quantities are calculated to the motor shaft using the
gear ratio N , which is constant in the case of a Harmonic Drive gear and posture-
dependent for the linear drives of the knee and ankle joints, cf. Eq. (2.9) on page 61.
τm is composed of the motor shaft acceleration torque resulting from the motor
inertia Jm and acceleration ω̇m . The latter is calculated from the load acceleration,
ω̇m = N ω̇l . τl is the load torque at the gear output. The friction torque τf is
composed of the motor and load-side friction, τf ,m and τf ,l , respectively, and sgn(·)
denotes the sign of the respective value.
Motor Selection Criteria

Motor selection is done on basis of the fundamental design data, Section 3.3.1,
and appropriate safety factors fi to account for task uncertainty2 . Three main
criteria have to be fulfilled:
1. The continuous stall torque rating of the motor has to be greater than or
equal the effective torque of the load profile to avoid overheating
2. The peak torque rating of the motor has to be greater than the maximum
required torque to avoid demagnetization of the rotor magnets
3. The speed limit of the motor has to be above the maximum speed required
by the load profile
The root mean square of the required motor torque during interval T is given by
s Z
1 2
τm,avg = τm dt . (3.2)
T T
Using Eq. (3.1) with Eq. (3.2) and the continuous stall torque rating τc of the motor,
the first criterion is obtained under consideration of the continuous load safety
factor fc :
s
Z 2
1 τ
τc ≥ fc Jm N ω̇l + l + τf dt . (3.3)
T T N
2 All safety factors fi for drivetrain dimensioning range from 1.5 to 2.

In order to fulfill the second criterion, the maximum torque during the load cycle
must be below the motor peak torque τp , including the static load safety factor fs :
τl
τp > fs max Jm N ω̇l + + τf .
(3.4)
N
Finally, the maximum permissible speed of the motor ωm,max must be greater than
the maximum speed during the load profile, using the speed safety factor fωm :
ωm,max > fωm max |N ωl | . (3.5)
If several motors match Eqs. (3.3) to (3.5), the motor constant Km , see Eq. (2.4)
on page 41, is used as an indicator for the power-to-torque ratio. In this case, Km
must be considered together with a motor’s thermal capability to determine its
true efficiency and capacity.
Derivation of the PMSM Speed/Torque Curve

The survey in Eq. (2.4) showed the superior power density of PMSM resulting
from the brushless design and thermal capabilities. For proper motor dimen-
sioning with appropriate reserve capacities it is necessary to examine the PMSM
speed/torque curve. Assuming identical phase voltages for all motors, the torque-
speed characteristic depends on the machine winding and is limited by a voltage-
dependent torque-speed boundary and a temperature boundary which depends
on the thermal environment of the motor.
For a particular phase voltage U there is a linear relationship between motor
speed ωm and torque τm , representing the torque-speed boundary of the motor in
steady state:
U R
ωm = − 2 τm . (3.6)
Kτ Kτ
Here Kτ and R are the torque constant and winding resistance, respectively.
Although motor losses are neglected in Eq. (3.6), they must be considered in a
second step to determine the continuous operation area of the motor. The two
main sources of motor dissipation are electrical losses in the stator windings
and magnetic losses in the steel laminations of the stator. Electrical losses are
expressed as ohmic dissipation PR = I 2 R. Magnetic losses, or core losses, Pc
comprise eddy currents and hysteresis losses. They are typically expressed as an
exponential function of the motor speed ωm and core losses Pc0 measured at a
speed of 1,000 1/min. The core losses of the chosen frameless motor family from
Parker Bayside [141] are rated as
Pc = Pc0 ωm/2 .
3
(3.7)
The maximum allowed dissipation Pm,max of the motor is limited by the maximum
allowed temperature ϑm,max of the winding isolation3 and the ambient temperature
3 Typically, the stator temperature ϑm must not exceed 155 ℃.

ϑ∞ in which the motor is operating:
ϑm,max − ϑ∞
Pm,max = . (3.8)
Rth,m
Here Rth,m is the thermal resistance between the windings and the ambient. Rth,m
depends on the size and mechanical design of the motor and the thermal properties
of the surrounding structure, and can be reduced by cooling fins, or forced air or
water cooling. Subtracting Eq. (3.7) from Eq. (3.8) yields the maximum allowable
electrical losses. Using the torque-current relationship, τm = Kτ I, we finally get
the thermal boundary of the torque-speed curve:
" !#2/3
1 ϑm,max − ϑ∞ R 2
ωm = − 2 τm . (3.9)
Pc0 Rth,m Kτ
Obviously, the output is limited by the temperature rise in the stator coils. The
load rating is thus greatly enhanced if the motor is tightly integrated into the
robot structure which can serve as a high-capacity heat sink. Note that the motor
temperature also affects the winding resistance and magnetic characteristic values.
As a consequence, the torque-speed curve, Eq. (3.6), becomes steeper and the
peak torque τp is reduced. Winding resistance increases linearly according to the
thermal resistance coefficient αCu for copper4 , and the winding resistance at motor
temperature ϑm becomes R = R∞ [1 + αCu (ϑm − ϑ∞ )]. Hence, winding resistance
increases up to 1.5 times the nominal value R∞ for the maximum permissible
motor temperature ϑm,max . Practical experiments with the robot Johnnie have
shown that the motors reach their thermal steady-state after 20 to 30 minutes of
continuous operation [109]. As the dynamics simulation data usually account for
less than one minute, the transient heating of the motors is not considered and a
constant value for the winding resistance is used instead. It is therefore proposed
to create a safety margin by using the value for ϑm,max throughout the motor sizing
calculations.
Using Eqs. (3.6) and (3.9), the continuous speed/torque curve can be plotted as
shown in Figure 3.3. The areas of constant and intermittent operation are shown
in ivory and orange, respectively. Since servo motors usually do not operate in
continuous duty, the peak torque τp during acceleration can be much larger than
the continuous stall torque: if the effective dissipation
R 2
Pm,avg = τm,avg , (3.10)
Kτ2
as calculated from Eq. (3.2), is within the continuous operation limits, the motor
can be overloaded intermittently. Typical peak torque multiplication ratios for
rare-earth-magnet based motors range from 3 to 5. The limitation is mainly due
to saturation effects in the loaded motor which cause a decrease in Kτ to 85 % to
90 % of its nominal value [141].
4 αCu = 3.92 · 10−3 1/K

(a) Torque-speed diagram (b) Motor dissipation

8000
1600
7000
1400
6000
1200
5000
Losses [W]
1000
4000
800
3000 600 Pm,avg = 493.6 W
2000 400
Pm,avg = 193.4 W
1000 200
b b
0 0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17
Motor torque |τm | [Nm] Time t [s]
Continuous operation Gait pattern (N =80)

Intermittent operation b Operating point (N =50)
Gait pattern (N =50) b Operating point (N =80)
Figure 3.3: Left, torque-speed diagram of the hip joint flexion/extension motor for a
walking speed of 5.55 km/h; right, motor power dissipation for different reduction
ratios.
Selection Criteria of Transmission Components

The drive mechanism of a joint is composed of the main transmission which inter-
faces to the links of the robot and in some cases auxiliary transmissions to connect
the electric motor to the main transmission. For the main transmissions, Harmonic
Drive gears [59] and planetary roller screws [78] are used. Synchronous belt drives
[24], angular gears [55] and coupler links serve as auxiliary transmissions. The
dimensioning of other drivetrain components, such as bearings [79] and couplings
[159, 160] follows the guidelines of the respective manufacturer. Again, safety
factors in the range from 1.5 to 2 account for task uncertainty.
Although the design and working principle of Harmonic Drive gears and plan-
etary roller screws are apparently different, there are some similarities of the
selection procedure which are addressed in the following. As the load torque τl
and speed ωl are varying, equivalent operating values are determined for both.
The average load speed is given by
Z
1
ωl,avg = |ω | dt . (3.11)
T T l
The average load torque is calculated as the root mean cube of the load torque τl ,
weighted by the load speed:
v
tR
3
|τ (t)3 ωl | dt
T Rl
τl,avg = . (3.12)
T
|ω l | dt
The transmission ratio N sets the torque, speed and inertia relationship of the
servo drivetrain to the link. Moreover, it affects the motor dissipation. Figure 3.3
illustrates the selection of a suitable motor-gear combination. For example, the

hip joint flexion/extension axis5 , shows that the torque-speed characteristics of
PMSM allow to decrease the gear reduction ratio N , while the efficiency, denoted
by the ratio of the load torque to the motor shaft acceleration torque, increases.
Figure 3.3(a) compares the torque-speed requirements for the reference walking
pattern for reduction ratios of N = 50 and N = 80. The shaded areas correspond
to Eqs. (3.6) and (3.9) and represent the continuous and intermittent operation
area of the motor6 , respectively. For N = 80, the higher motor speed increases the
acceleration torque of the motor shaft such that a larger motor must be selected.
Conversely, for N = 50 a less powerful and more lightweight motor can be chosen.
Likewise, the operating point for N = 50 denoted by the RMS torque, Eq. (3.2),
and the mean rotational speed, calculated analogous to Eq. (3.11), is within the
continuous operation limits of the motor. Thus, the motor is suitable to drive
the joint and can be overloaded intermittently. Conversely, the operating point
for N = 80 is outside the continuous operation limits and would necessitate a
larger motor. The effect of lower reduction ratios can also be seen from the motor
dissipation in Figure 3.3(b). The average motor dissipation Pm,avg , Eq. (3.10), is
rated at 193.4 W for N = 50, whereas Pm,avg increases significantly to 493.6 W
(=255 %) for N = 80.
Elastodynamic Analysis of the Drive Mechanism

While the entire drive mechanism was considered rigid and free from backlash for
the torque-based dimensioning, Eq. (3.1), the finite drivetrain stiffness is a crucial
factor in meeting the stiffness requirements of the mechanical structure: because
transmission elements, such as gears, belt drives and couplings introduce compli-
ance between the motor and the load, they bring up the problem of mechanical
resonance which may limit the bandwidth of the mechatronic servo system and
provoke weakly damped oscillations of the link. The aim of the elastodynamic
analysis of the drive mechanisms is therefore to
– eliminate or reduce resonances in the drive mechanism caused by the finite
stiffness of transmission components;
– achieve a high enough bandwidth and fast response time of the speed control
loop;
– minimize power dissipation.
While the stiffness of planetary roller screws is sufficiently high, the distinct com-
pliance of synchronous belt drives and Harmonic Drive gears must be considered
in the elastodynamic analysis of the drive mechanism. Backlash also affects the
bandwidth of a servo system [10] but can be neglected for Harmonic Drive gears
and pre-loaded planetary roller screws.
Although the drivetrains of the robot are rather complex structures, the major
effects of the compliant motor-load coupling can be understood by basic analytical
considerations. Figure 3.4a shows the reduced model of a drivetrain with elastic
gearbox. The motor and load inertia, Jm and Jl , are coupled by a backlash-free
5 The complete load profile of the hip joint is given in Appendix D
6 The relevant technical data of the motors are given in Appendix C
Gearbox (N , kϕ )
ϕm , τm ∗ = Nτ
τm m
ϕl τl ∗ = ϕ /N
ϕm kϕ ϕl τl
m
Motor (Jm ) Load (Jl ) ∗ = N 2J )

Motor (Jm Load (Jl )
m
(a) Drivetrain with elastic gearbox (b) Two-mass system with flexible coupling
Figure 3.4: Equivalent model of a compliant drive mechanism
gearbox with constant ratio N and torsional stiffness kϕ . The motor and load-
side masses of the gearbox are lumped in Jm and Jl , respectively. For the sake of
simplicity and because it is a worst case assumption, damping is neglected. The
motor torque τm and load torque τl are active forces acting upon the masses Jm
and Jl , respectively. Reducing this model by reflecting the motor quantities to
the load side, leads to the unconstrained two-mass system with flexible coupling
shown in Figure 3.4b. The angle ϕm ∗ on the motor side differs from the angle ϕ on
l
the load side. The system is described by two ordinary differential equations for
the motor and load side, respectively:
∗ ∗ ∗ ∗
Jm ω̇m + kϕ (ϕm − ϕl ) = τm , (3.13)
∗
Jl ω̇l + kϕ (ϕl − ϕm ) = τl . (3.14)
Here (·)∗ denotes motor quantities calculated to the load side. Note that the
resultant load inertia Jl is composed of several link inertias. Hence Jl is posture-
dependent for most drives and the elastodynamic properties must be analyzed for
a set of typical postures.
Using Eqs. (3.13) and (3.14), the transfer functions of the motor and load speed
are determined. The transfer functions Eqs. (3.15) and (3.16) describe the motor
and load response, respectively:
k
∗ (s)
ωm s2 + Jϕl
∗ = ∗ (3.15)
τm (s) s s2 + kϕ Jl +J∗m
J Jm l
kϕ
ωl (s) ∗J
Jm l
∗ = ∗ (3.16)
τm (s) s s2 + kϕ Jl +J∗m
J Jm l
Here s is the complex variable. The natural frequency and antiresonance of the
mechanical system, ωr and ωar , correspond to the poles and zeros of the transfer
functions, respectively, and are written as
s s s
∗
kϕ kϕ (Jl + Jm ) kϕ (Jl + Jm N 2 )
ωar = , ωr = ∗ = . (3.17)
Jl Jl Jm Jl Jm N 2
From Eq. (3.17) it can be seen that ωar and ωr are determined by the motor and
load inertia, as well as the gear ratio and stiffness. The antiresonance frequency,
ωar , is the frequency where an oscillation node coincides with the point of force
application [32] and can be taken as a natural frequency at the load side. A forced
torsional oscillation of the load side at ωar may cause the load side to oscillate at ωar
while the motor remains at standstill or nearly standstill (ωm ≈ 0, τm ≈ 0), and the
servo actuator may fail to respond to desired actions or compensate disturbances.
As an empirical value recommended by the gear manufacturer, far = ωar/2π should
be above 15 Hz for industrial robots and standard applications in mechanical
engineering to avoid such effects [59]. During operation, the resonance speed on
the motor side resulting from far should be passed rapidly.
Load-to-Motor Inertia Ratio

From the control systems perspective, minimal motor inertia Jm yields faster
accelerations and better transient responses due to the high torque-to-inertia ratio.
This only holds true for the (unrealistic) assumption of a rigid motor-load coupling.
In reality the load response does not correspond with that of the motor because
of the finite stiffness of the motor-load coupling, cf. Eqs. (3.13) and (3.14). In
this case the load-to-motor inertia ratio is fundamental in limiting the bandwidth
of the servo system and, thus, an important parameter for drivetrain sizing and
indicator of possible stability problems. The load-to-motor inertia ratio Rj is
derived from the ratio of the natural frequencies, Eq. (3.17):
ωr q Jl Jl
= 1 + Rj , where Rj = ∗ = . (3.18)
ωar Jm Jm N 2
Servo systems with low ratio Rj show good disturbance rejection and response
times. In order to prevent overshooting and oscillation at larger ratios (Rj > 5), the
control gains must be reduced which extends the settling time. Another solution is
to use control schemes which take the compliance of the motor-load coupling into
account, see for example [29, 166]. According to a closed-loop analysis by Zhang
and Furusho [209], the load-to-motor inertia ratio in high-response applications
should be 1 ≤ Rj ≤ 4. Finally, Rj affects power dissipation of the servo actuator
because the motor acceleration torque increases with larger inertia ratios which
goes along with higher power dissipation.
Practical Considerations
Rivin [155] addresses fundamental issues that must be considered when analyzing
the stiffness and damping of power transmission systems and servomechanisms.
The antiresonance and resonance frequencies, Eq. (3.15), greatly affect the
bandwidth of the speed control loop and the response time of the servo actuator
[10]. In order to avoid resonance problems and to achieve faster tracking and
closed-loop response, it is essential to increase the natural frequencies of the
drivetrain without increasing its mass. To this end, the most effective measures to
improve or avoid resonance problems are
– reducing the load inertia,

– choosing a higher transmission ratio,
– increasing the torsional stiffness of the drive mechanism, or
– using a motor with higher (peak) torque-to-inertia ratio.
Obviously, consistent lightweight design of the link structure and an improved
mass distribution—of individual links and the kinematic chain as a whole—is the
best way to reduce the load inertia Jl . If the lightweight design is fully exhausted,
higher gear ratio or lower pitch for roller screws can be chosen to adjust the load-
to-motor inertia ratio Rj . The coupling stiffness of the motor to the load is another
critical issue in the servo drivetrain. Higher torsional stiffness of the motor-load
coupling leads to higher natural frequencies of the overall system. As an example,
the size of the Harmonic Drive gear unit of the hip joint flexion/extension axis is
not only due to the higher torque loading, rather, it increases the stiffness of the
motor-load coupling. Motor dynamics should rather be improved by choosing a
motor with a higher dynamic factor (i. e., higher peak-torque-to-inertia ratio, cf.
Eq. (2.4)) than by increasing the motor inertia. A higher dynamic factor accounts
for better acceleration capabilities, thus, the motor can tolerate higher Rj with
good response.
3.3.3 Dimensioning of Structural Components

The link structure of Johnnie accounts for nearly 43 % of the overall weight [46]
which indicates a certain weight saving potential. The primary objective of struc-
tural dimensioning is therefore to keep the mass of structural components as
low as possible while the static and dynamic stiffness should be high enough to
minimize distortion and structural resonances. On the other hand, as discussed
in Section 3.2, fully exploiting the weight saving potential is not feasible due to
uncertainty of loading.
General Design Approach
Two different development methodologies are employed during the concept design
stage:the empirical design approach relies on the designer’s experience and intuition
and is the most appropriate way of developing most part geometries. For major
structural components with complex multi-axial stress conditions and/or strict
geometric constraints, topology optimization is used to generate concept design
proposals. Figure 3.5 compares the workflow of both approaches.
The empirical design shown in the left-hand branch is a very efficient trial-
and-error approach which allows to develop well thought-out designs for most
components, taking into account low mass and ideal force flux. Starting from a
logical point, the design of single parts and/or assemblies is iteratively refined
using the built-in finite element analysis (FEA) modules of CATIA7 for structural
analysis and strength assessment. Constraints are defined to resemble the actual
installation of the component or assembly. Depending on the diversity of load
7 CATIA Generative Part Structural Analysis (GPS) and CATIA Generative Assembly Struc-
tural Analysis (GAS) are used for the analysis of single parts and assemblies, respectively.
Both GPS and GAS allow the consistent mapping of native CAD models to finite element
models and integrate pre-processing, solving and post-processing capabilities.
CAD modeling CAD draft / Mockup
Empirical design Set up finite-element model Set up finite-element model
Topology optimization
FEA / Strength assessment Set up optimization problem
Interpretation of result Numerical optimization
OK NOK Goal achievement?

Modification or
yes no
re-design
Modification of
element density
Final CAD design Design proposal
Figure 3.5: Comparison of empirical design to the automated process based on topol-
ogy optimization
conditions, one or more load cases are analyzed using the constraint forces and
moments calculated with the dynamics simulation (Section 3.3.1). This approach
is particularly useful for smaller components and/or components where the
weight saving potential is mainly limited due to manufacturing constraints and
capabilities (e. g., minimum wall thickness or cutter radii).
If a component has various complex load conditions, strict geometric con-
straints and/or other design limitations, the empirical design approach is not
always straightforward. Rather, it is difficult to find a satisfactory solution which
meets the requirements. Here the FEM-based topology optimization method [61]
aids the designer in achieving an effective lightweight design before modeling the
actual part geometry. The additional workflow is shown in the right-hand branch
of Figure 3.5. After the allowable designed space has been defined as a solid body
or shell structure in CAD, the finite-element model is generated including all
loads and constraints required to resemble the installation of the component or
assembly as realistic as possible. Then the optimization problem is formulated,
using mass or stiffness targets as possible goal functions. Optimization constraints
include the maximum allowable deflection, minimum wall thickness, symme-
try and manufacturing process parameters. Also, dedicated areas are reserved
that must not be modified, e. g., bearing seating surfaces or connecting flanges.
The design proposal is then determined iteratively without user interaction by
numerical optimization. The result is mathematically optimal and superior to
conventional designs, provided that the whole optimization model has been set
up properly. Several load cases must be considered for optimization in order to
obtain results which apply to typical, realistic scenarios. However, the implemen-
tation and verification of an optimization model are very time-consuming tasks.
Also, transferring the design proposal into an actual part geometry is very costly
M
F
z
x y
(a) Mockup (b) Optimization model (c) Optimization result (d) Part design
Figure 3.6: Topology optimization of the shank
because special attention has to be paid to manufacturability and assembly. The

use of topology optimization must therefore be carefully balanced regarding the
cost-benefit ratio and is thus limited to very few and complex parts. Another
problem relates to the uncertainty of loading: as the loads can vary significantly
when testing different walking patterns and control strategies, appropriate safety
margins must be taken into account in the dimensioning of the robot structure.
Topology optimization can thus only give an indication for the geometry of the
actual part under typical load conditions.
The main bodies of the pelvis [168], thigh and shank segments have a very
complex Monocoque-like structure (cf. Section 3.10), suggesting to create design
proposals by topology optimization8 . The optimization process is illustrated
in Figure 3.6 using the example of the shank that connects the one-DoF knee
joint and the two-DoF ankle joint [112]. Since these joints are actuated by linear
mechanisms, the loads are transferred not only at the joint flanges but also at the
several hinging points and bearings of the drive mechanisms.
In a first step the CAD mockup is generated which resembles the available
designed space of the shank, Figure 3.6a. To this end, an interference analysis is
conducted between the mockup solid and the operating volumes of moving parts,
i. e., the drive mechanisms of knee and ankle joint and the adjacent links. For
example, the space reservation for the linear actuator of the knee joint was done
through the swept volume of the moving parts created with CATIA’s DMU mod-
ule. The mockup geometry should already consider the intended manufacturing
8 The commercial software HyperWorks/OptiStruct from Altair Engineering was used.
M M
F F
z z
x y x y
(a) FE model (b) FE mesh (c) Deflection (d) von Mises stress
Figure 3.7: FEM-based strength assessment of the shank
process and capabilities. The optimization model shown in Figure 3.6b is built
upon the mockup geometry. Since realistic results can only be achieved if the load
transfer by the drive mechanisms of the knee and ankle joint is taken into account,
the shank and the knee and ankle joint drives are modeled as elastic bending
beams. The constraint forces and moments are applied to the hip joint, while the
ankle joint and coupler links are cardanically supported to the environment. The
visualization of the results allows to locate the load step with the most significant
contribution to the maximum load case. After the optimization results have been
prepared by calculating an iso-surface of the material distribution, smoothing and
data reduction, they are imported in the CAD system. The optimization result
is shown in Figure 3.6c: the left-hand plot resembles an iso-plot of the element
density. The right-hand vector plot shows the principal stresses, indicating the
force flux. Finally, the part geometry shown in Figure 3.6d is developed based on
the optimization result.
Strength assessment
Iterative stress analyses are applied throughout the design process as it allows the
rather fast and simple evaluation and comparison of different design approaches,
see Figure 3.5. After completion, highly stressed components and sub-assemblies
Mode 1 (317 Hz): side-to-side bending Mode 2 (360 Hz): upward bending
Mode 3 (902 Hz): torsion Mode 4 (997 Hz):

side-to-side bending & torsion
Mode 5 (1204 Hz): upward bending
Figure 3.8: Natural mode shapes of the shank base body. Scaling is omitted since the
calculated deflections are only of qualitative nature. The red and blue areas indicate
zero and maximum deflection, respectively.
are assessed by finite element analyses (FEA) for a final strength and fatigue
evaluation [54]. The constraint forces and moments are again calculated with
the comprehensive robot simulation (see Section 3.3.1). Based on these dynamic
load profiles, different load cases are identified, e. g., maximum force, maximum
moment and average loading, and safety factors account for load uncertainty. The
material data are taken from different sources, such as manufacturer’s data sheets
or the standard work by Haibach [54]. Because of high manufacturing effort
and costs, major link structures and some other critical components are designed
conservatively to avoid failure. According to the infinite lifetime concept, stresses
are kept below the threshold of fatigue limit.
The FEA of the shank sub-assembly is shown in Figure 3.7 to exemplify the
workflow of strength assessment. In order to obtain a realistic and accurate
finite element model, Figure 3.7a, it is not sufficient to consider solely the main
body of the shank. Rather, a sub-assembly must be built consisting of the shank
base body, knee joint flanges and linear carriages of the ankle joint drives (cf.
Section 3.8.3). Moreover, the surrounding construction, including the thigh, knee
joint actuator and coupler links of the ankle joint are modeled as elastic bending
beams. The 1-D and 3-D meshes are connected by assembly constraints that
resemble fastened, sliding, rigid, or smooth connection properties. Since influence
on CATIA’s automated meshing algorithms is limited, it is essential to perform
iterative mesh refinement to obtain reliable results, Figure 3.7b. Lastly, Figure 3.7c
and d illustrate deflection and von Mises stress of the component.
Elastodynamic Analysis
The dynamic stiffness of the leg links determines to a far extent the positioning
accuracy of the feet. Structural vibrations can let the foot-ground contact occur too
early and may, thus, easily destabilize the robot. Therefore, the resonant behavior
of the major leg links is analyzed numerically with the FEM to determine the
natural mode shapes and frequencies of the structure during free vibration. Nu-
merical modal analysis supplies information about deflection modes and natural
frequencies which are present in the robot structure. The results can help tracing
poorly designed elements and structural weak points during the design stage. For
example, Figure 3.8 visualizes the natural frequencies and corresponding mode
shapes of the shank main body. They are basically characterized as side-to-side
bending, upward bending, torsion, and combinations thereof. The figure only
depicts the first five frequencies which are typically the most prominent modes at
which the shank will vibrate, dominating all the higher frequency modes.
3.4 Modular Joint Design

The detailed weight analysis by Gienger [46] revealed that the servo actuators
make the second largest part the total weight—22.7 % account for the motors and
another 7.9 % for the gears. Along with the constraints of human dimensions,
the weight restrictions exclude the use of off-the-shelf servo actuators. Hence
the development of lightweight servo actuators that combine precision gearing
with a high-dynamic servo motor and sensors into a compact package is cru-
cial. This section presents the design of the revolute joints based on permanent
magnet synchronous motors and Harmonic Drive gears that were discussed in
Section 2.5.1.
3.4.1 Motivation for Modularity and its Limitations

Modularity as a design goal has been extensively investigated in the engineering
literature [27, 138]. Regarding robotics, modularity is a rather new but suitable
approach to designing complex robotic systems. Examples of modular robots built
from a set of standard components include research prototypes [67, 120], as well
as commercially available systems9 .
Besides well-known design and manufacturing benefits, the principal advan-
tages of modular design include reduced time-to-repair and spares inventory.
Moreover, reconfiguration of the robot (to modify certain components according
to experimental results, or match different operative requirements), performance
upgrades of the actuators, and future extensions (additional DoFs, sensors, etc.)
are simplified and have minor effects on the rest of the system. For example, the
arms currently have a quite simple structure (cf. Section 2.2.4), tailored to com-
pensate the angular momentum effectively. From a hardware point of view, the
replacement by full arms and hands for manipulation is simplified by a modular
architecture. The major disadvantages of modular robots are the higher overall
9 For example, the Schunk PowerCube system, available at: http://www.schunk.com/
3.4 Modular Joint Design 97
mass, lower dynamic performance and reduced stiffness. The overall mass is gen-
erally higher because of the larger number of parts and their connections and goes
along with lower dynamic performance. Moreover, since all servo actuators are
located close to the corresponding joints, the inertia of modular serial-link robots
is higher than at comparable non-modular systems. If properly designed, the
latter have an improved mass distribution since the spatial separation between the
actuators and joints is possible. Finally, the stiffness of modular serial-link robots
is reduced due to the higher part count, the presence of numerous connections
and, frequently, unfavorable force flux resulting from the standardized flanged
connections and joint bearing structures.
Therefore, a fully modular robot architecture, where all functional unit are
independent from each other, is not useful for a high-dynamic robot and con-
tradictory to the fundamental design goals developed in Section 2.4. Rather, a
limited modular approach with actuators built on the unit construction principle
is proposed, while the main structure is non-modular. All rotary drives have the
identical structure and comprise a joint, the corresponding actuator (motor and
reduction gear) and a servo drive unit that also incorporates a low-level motion
controller (see Section 5.2). The sizes of the motors and gears are adapted to
the individual power requirements. Many parts are standardized for all drives,
only some housings and connecting flanges are customized for optimal force flux.
Moreover, the linear actuators of the knee and ankle joints (cf. Sections 3.8.2
and 3.8.3) employ the modular design features to a far extent. This approach
turned out to be the most reasonable and economic way of realizing the robot at
minimal mass while taking into account ease of manufacturing.
Although the main structure is non-modular, standardized flange connections
of many actuators (such as the shoulder and hip joints) and, particularly, the
separation of actuators and structural components yield certain modularity on a
higher level, allowing to easily exchange the arms and legs, as well as the upper
body. Conversely, a rather low level of modularity with joint-specific connections
is applied to the legs in order to achieve a particularly advantageous force flux,
high stiffness and low mass.
3.4.2 Revolute Joints based on Harmonic Drive Gears

Highly integrated joint units with maximum power density can only be realized,
if the latest technologies in the field of electrical drives, gears and sensors are used.
As stated in Section 2.5.1, the main reasons to choose frameless PMSM over BDCM
are their torque and speed bandwidths, the greater power density and robustness.
The fusion of motor, gear and sensors into highly integrated, mechatronic joint
units has several advantages for the overall system:
– High velocity range at good dynamic performance
– High power density and efficiency
– Comparatively small designed space per servo actuator
– Reliability due to brushless design of the motors and, in conjunction with de-
centralized motor control, the ability to self-monitoring and fault diagnosis
(cf. Section 5.2)
A
B
C
D E
F
G
0 50 100 [mm]
A Hip joint (flexion/extension) E Shoulder joint (flexion/extension)

B Hip joint (adduction/abduction) F Elbow joint (flexion/extension),
C Hip joint (internal/external rotation) Shoulder joint (adduction/abduction)
D Pelvis joint (adduction/abduction), G Toe joint (flexion/extension)
Pelvis joint (internal/external rotation)
Figure 3.9: Modular design of revolute joints based on Harmonic Drive component sets
(drawn to scale)
Except for the camera head described in Section 3.5 where the low power require-
ments allow to use off-the-shelf servo actuators, a total of 16 DoFs are actuated
by seven different rotary drives based on Harmonic Drive component sets. All
gears are customized lightweight versions with a T-shaped Circular Spline which
is, in the author’s opinion, the best tradeoff between mass and load rating. The
solid Wave Generator plugs are made from either aluminum or steel and modified
for low mass and inertia. Most drive units employ HFUC series component sets
[59], except for the hip joint flexion/extension axes. Here, CSG series component
sets [59] are used as they feature a significant increase in torque capacity in com-
parison to the HFUC series gears. The toe joint actuators employ off-the-shelf
CPL series lightweight component sets [57]. Figure 3.9 gives an overview of the
drives, starting from the high-performance drive ‘A’ used in the hip joint flex-
ion/extension axis, to the lightweight toe joint actuator ‘G’. Note that the type ‘B’
and ‘C’ actuators share the identical motor housing and differ only in the output
bearing. Type ‘D’ and ‘E’ differ only in the stack length of the motor, so that drive
‘E’ is omitted in the figure. The key technical data are listed in Appendix C.
Basically, two joint principles are employed that differ in the output flange
bearing. The first type uses a four-point contact ball bearing and an additional
support bearing. This type is applied to joints where a single bearing at the output
flange would impair the force flux and unacceptably decrease the tilting rigidity
of the joint, for instance the hip joint adduction/abduction and flexion/extension
axes. The second type employs a thin-ring crossed roller bearing at the output.
Although a single bearing is used, tilting rigidity and load rating are maintained,
space restrictions can be met easier and the adjacent construction is simplified
and more compact.
The design and mode of operation of the servo actuator with four-point con-
tact ball bearing are illustrated by Figure 3.10 which shows the hip joint flex-
ion/extension drive (type ‘A’). For optimal heat transfer, the stator (1) of the
3.4 Modular Joint Design 99
23 2 8 13
22 15
21 20 1 Stator
2 Motor housing
3 Permanent magnet rotor
1 4 Motor shaft
5 Motor shaft bearing A
6 Motor shaft bearing B
10 7 Wave Generator
8 Circular Spline
9 Flexspline
10 Output flange
11 Clamping ring (Flexspline)
12 Gap seal
13 Four-point bearing
14 Sensor mounting ring
15 Incremental encoder
16 Sensor shaft
17 Absolute angular sensor
17 19 18 Elastomer coupling
18 14
19 Sensor bearing
20 Limit switch
12 16 12 7 21 Protective cover
5 4 11 9
3 6 22 Cable grommet
23 Support bearing
Figure 3.10: Harmonic Drive-based revolute joint with four-point contact ball bearing
and support bearing, here: hip joint flexion/extension drive (type ‘A’)
frameless PMSM is bonded into the motor housing (2) with a thermally conduc-
tive epoxy. The motor housing (2) has cooling fins for natural convection. The
permanent magnet rotor (3) is bonded onto the motor shaft (4) which rotates on
two deep groove ball bearings in a locating/non-locating arrangement. The locat-
ing bearing (5) is placed next to the sensor area to maintain the axial position of the
magnetic ring of incremental encoder (15), ensuring optimal output signals and
reducing interpolation errors. The non-locating bearing (6) is placed on the gear
side where axial displacement due to thermal expansion is not critical. The Wave
Generator (7) is flanged on the motor shaft (4). The Circular Spline (8) is fixed to
the motor housing (2), and the Flexspline (9) is attached to the output flange (10),
using the clamping ring (11). Hence, turning the motor shaft (4) moves the output
flange (10) with respect to the motor housing (2). The EKagrip [34] friction shim
(not shown), inserted between the Flexspline (9) and the output flange (10), greatly
increases the coefficient of friction so that the maximum torques can be safely
transmitted. The gap seal (12) keeps the clearance between the gear components
and the housing within the specifications for grease lubrication. A four-point
contact ball bearing (13) is employed at the output flange (10). It is complemented
by the additional support bearing (23), which is a radial contact ball bearing. The
inner ring of bearing (23) is clamped in position by the sensor mounting ring
(14). The bearings (13) and (23) are sealed thin-section bearings with imperial
measurements which are widely used in robotics and aerospace applications. They
are characterized by low mass and small cross sections which remain constant
as the bore diameter increases. All sensors (cf. Section 4.1) are arranged on the
B-side of the motor in a separated dustproof section of the housing, opposite to
the gear unit. The incremental encoder (15) is mounted on the motor shaft (4) and
provides angular position and velocity of the permanent magnet rotor (3) with
23 24 17 21
13 9 18 19
11 8 15
2 16 12 1 Stator
2 Motor housing
3 Permanent magnet rotor
4 Motor shaft
5 Motor shaft bearing A
6 Motor shaft bearing B
1 7 Wave Generator
10 8 Circular Spline
9 Flexspline
10 Output flange
11 Clamping ring (Flexspline)
12 Gap seal
13 Crossed roller bearing
14 Sensor mounting ring
15 Incremental encoder
16 Sensor shaft
18 Elastomer coupling
20 19 Sensor bearing
14 22 20 Limit switch
3 5
6 4 21 Protective cover
7 12 22 Cable grommet
23 Clamping ring (bearing)
24 Grease nipple
Figure 3.11: Harmonic Drive-based revolute joint with crossed roller bearing, here: hip
joint internal/external rotation drive (type ‘C’)
respect to the stator (1). The flange of the absolute angular sensor (17) is fixed to
the gear housing using the mounting ring (14). It is connected to the gear output
(10) by the sensor shaft (16), enabling the direct measurement of the output angle
without obstructing the connection between the gear output and adjacent con-
struction. The sensor shaft (16) is free from moment and radial loading to ensure
an accurate transmission of the output angle. The groove on its output flange side
acts as an elastic flexure hinge. In conjunction with the backlash-free miniature
elastomer coupling (18), the sensor shaft (16) is angularly flexible to compensate
for radial and axial misalignment. The coupling (18) connects to the disk of the
angular sensor (17), which is supported by the sensor bearing (19). A double
row angular contact ball bearing is used as it takes up less axial space than two
single row bearings. The support configuration of the sensor shaft (16) and sensor
disk are thus statically determined. The short sensor shaft allows to maintain the
correct working distance between the disk and the measuring module of sensor
(17). Finally, the photoelectric fork sensor (20) limits the angular workspace of
the joint, using a cam fixed to the sensor shaft (16). The housing of the rotary
drive is divided into three areas: Harmonic Drive gear, electric motor and sensors
are installed in different areas, separated by the bearing flanges. Accordingly, all
bearings are equipped with seals, but gap seals are preferred over lip seals due to
reduced friction. In addition, the gap seals (12) prevent the grease from ingress
into the motor and sensor areas. Protection of the sensors against ingress of solid
foreign objects and dust is ensured by the cover (21) and the cable grommets (22).
The second type of rotary servo actuator shown in Figure 3.11 is applied to the
internal/external rotation axis of the hip joint. The basic structure is identical
to the previous actuator but this type actuator employs a single thin-section
3.5 The Stereo Robotic Head 101
6 2 3 4 5 9
6 2 3 4 5 9
7
1
1
8
8
(a) Hip joint internal/external rotation (b) Hip joint flexion/extension
1 Motor shaft 4 Adjustment shims 7 Wave generator

2 Permanent magnet rotor 5 Retaining ring 8 Screws
3 Locating bearing 6 Non-locating bearing 9 Magnetic encoder ring
Figure 3.12: Low-inertia design of the motor shafts
crossed roller bearing (13) at its output flange. crossed roller bearings are typically
supplied with pre-load and can receive loads in all directions, including radial,
axial and moment loads. Due to the line contact between the cylindrical rolling
members and the bearing rings, these bearings achieve high tilting rigidity and
load ratings, eliminating the need for an additional support bearing. The bearing
outer ring is easily fixed to the housing by the clamping ring (23), while the inner
ring is fixed by the adjacent construction. During assembly of the drive unit,
the clamping ring (23) is fixed only by a few small screws. It is then tightened
tension-free and with the prescribed number of screws and tightening torque,
when the drive unit is assembled in the adjacent construction. The grease nipple
(24) eases maintenance of the crossed roller bearing (13).
Figure 3.12 illustrates the low-inertia design of the motor shafts. Since motor
shaft inertia is dominated by the inertias of the permanent magnet rotor (2) and
Wave Generator (7), the latter is customized for low inertia and mass. The solid
Wave Generator plug is made from aluminum or steel and omits the Oldham
coupling of the standard versions. The permanent magnet rotor (2) is bonded onto
motor shaft (1) while the Wave Generator (7) is screwed to it. The motor shaft (1)
is made from aluminum and shows extremely thin cross sections. It is supported
by locating bearing (3), which is fixed in position by adjustment shims (4) and
retaining ring (5), and non-locating bearing (6). Both bearings are thin-section
ball bearings. The magnetic encoder ring (9) is clamped onto the motor shaft.
3.5 The Stereo Robotic Head

The head of the robot holds a high-resolution stereo camera pair. Its mechanics
presented in Figure 3.13 comprise a two-DoF pan-tilt unit whose torque/speed
capabilities allows smooth pursuit movements, as well as the generation of human-
5
3
4
4 1 Head mount 17 Support bearing
2 Pan segment 18 Coupler link
3 Tilt segment 19 Spherical plain bearing
30 4 Camera 20 Tilt bearing
2 5 Camera receptacle 21 Locknut
22 13 6 Pan motor 22 Limit switch
8 7 Pan gearbox 23 Servo actuator
7 8 End stop 24 Bellows coupling
6 9 9 Limit switch 25 Output bearing
11 10 DSP module 26 Lever arm
11 Motor controller 27 Coupler link
12 12 DC/DC converter 28 Vergence axis
1 13 Cable gland 29 Plain bearing
14 Servo actuator 30 Limit switch
10 15 Torque limiter 31 Camera spacer
16 Lever arm
(a) Overview
21 20 21 3
24 26
25
3 27
5 19
5 4
19 29
14
18 28
15 4
19 29
17
16
23
(b) Tilt drive mechanism (c) Vergence drive mechanism
5 4 5 31 4
133 mm 240 m
m
(min.) (max.)
(d) Stereo basis adjustment
Figure 3.13: Mechanical design of the three-DoF camera head

3.5 The Stereo Robotic Head 103
like saccadic camera motions. In addition, a one-DoF mechanism is implemented

for camera vergence adjustment, enabling the robot to fixate a visual target with
both cameras. The camera stereo basis is manually adjustable within a wide range,
and the mechanical camera interface allows to exchange the cameras very easily
and precisely. The total mass is 2.4 kg, including the cameras and lenses (0.31 kg
each).
The head mount (1) shown in Figure 3.13a connects to the upper body. It
is composed of two connecting flanges, bonded and riveted to a thin-walled
aluminum tube. The upper flange of head mount (1) accepts the pan servo motor
(6) and the reduction gear (7). The latter is a very compact and lightweight, off-
the-shelf gearbox with integrated output bearing [58]. The angular workspace of
the pan segment (2) is limited by two end stops (8). Due to limited space, absolute
angular sensors are omitted in the camera head: the switches (9) are precision
switches with high repeatability, serving both as limit switches and reference
point signals. Likewise, two limit switches (22) for tilt motion are mounted on the
pan segment (2), and the limit switch (30) for vergence motion is mounted on the
tilt segment (3). The pan segment (2) carries the tilt segment (3) which holds the
cameras (4) with the receptacles (5). The mechanical camera interface simplifies
experiments with different types of cameras. Each camera (4) is positioned with
respect to the receptacle (5) by two locating pins (not shown), ensuring fast and
easy replacement with high repeatability. Finally, the head mount (1) holds all
the necessary electronics, consisting of DSP module (10), motor controller (11)
and DC/DC converter (12) for conversion of the motor voltage. Cable gland (13)
guides all cables coming from pan segment (3).
Figure 3.13b shows the tilt axis drivetrain. The Harmonic Drive servo actuator
(14) [60] is connected to the tilt segment (3) by a parallel-crank mechanism to
achieve a compact design: the lever arm (16) is mounted on the output shaft of
servo actuator (14), driving the tilt segment (3) via the coupler link (18). The
coupler link (18) incorporates spherical plain bearings (19) at its ends. Due to
the large distance between the output bearing of servo actuator (14) and lever
arm (16), the gear output is supported by bearing (17) for accommodation of
tilting moments. The tilt segment bearing (20) is composed of two thin-section
ball bearings in an adjusted O-arrangement and secured by the locknuts (21). As
the loads during an emergency can easily exceed the collision torque of actuator
(14), the output shaft is protected by the torque limiter (15) that provides precise
and fast overload protection [160]. If the preset torque limit has been reached,
the output of torque limiter (15) is disconnected from the input. The chosen
torque limiter is free from backlash in either direction of rotation and provides
single-position re-engagement, i. e., it will re-engage automatically 360° from the
original position.
Figure 3.13c shows the drive mechanism of the vergence axis, which is entirely
integrated into tilt segment (3). The mechanism is designed to ensure symmetric
adjustment of the camera angle using a single actuator. The camera vergence axes
(28) are driven by a Harmonic Drive miniature servo actuator (23) [60] and two
four-bar mechanisms that interface to the camera receptacles (5). The mechanisms
consist of a double lever arm (26) and two coupler links (27) of equal length. Lever
arm (26) is supported by two thin-section ball bearings (25) in an adjusted bearing
arrangement, the coupler links (27) employ spherical plain bearings (19). The
miniature bellows coupling (24) with integrated torque limiter [160] connects
the output of servo actuator (23) and the lever arm axis (26), reducing the risk
of damage in an emergency. Both vergence rotational axes (28) are supported by
plain bearings (29) with minimal clearance.
Finally, Figure 3.13d illustrates the manual adjustment of the camera stereo basis.
The stereo basis can be adjusted between 133 mm and 240 mm by interposing
camera spacers (31) between the receptacles (5) and the cameras (4).
3.6 The Torso and Pelvis

The upper body is the root segment of the robot multibody system that is used
for simulation and control of the robot. The pelvis is a highly loaded segment,
interfacing the legs and the upper body. The design of these segments is pre-
sented in Figures 3.14a and 3.14b, showing the front and rear view of the robot,
respectively. The figures show the camera head (2), described in the previous
section, the arms (3) and the pelvis segment (1). All components are mounted
on the trunk support structure which is shown in Figure 3.14c and described
later in this section. The arms (3) are mounted on the support structure by the
shoulder joint flexion drive (4) flange. The upper body also holds the inertial
measurement unit (7) which estimates the orientation and angular velocities, see
Section 4.4. Mounted on the back of the robot is the PC-based central control unit
(8) that is described in Section 5.1. The upper body and the pelvis segment (2)
hold several DSP modules and motor electronics (5), cf. Section 5.2, for the upper
body and hip joints. The DC/DC converters (6) convert the external small signal
voltage to supply all decentral controller modules and the central control unit.
The connector panel (9) on the back provides connectors and strain relief for all
external cables, i. e., the Ethernet connection, motor and small signal voltage and
the emergency stop cable. The panel also includes an emergency stop pushbutton
(10) to disable the motor supply in case of failure (cf. Section 5.4). The handles
(11) simplify handling of the robot during experiments and transport.
Figure 3.14c shows the trunk support structure which is basically designed as a
bonded and riveted pipe frame structure. Hence the cavities of the main (13) and
shoulder girder (14) could be foamed with polyurethane for improved vibration
damping if necessary. All attached parts shown in Figures 3.14a and 3.14b are
mounted on the support structure using pipe clamps or blind nuts. The latter are
a very effective solution when load-bearing threads are required in thin-walled
parts. The support structure is flanged on the pelvis internal/external rotation
joint (not shown) by the connecting flange (12). The main girder (13) extends from
the pelvis joint flange (12) up to the shoulder joint level where it merges into the
shoulder girder (14) and the shoulder joint flanges (16). The shoulder transverse
tubes (15) and struts (19) provide additional load-moment support and reinforce
the connection of the shoulder joint flanges (16). The support structure provides a
connecting flange to the head mount (17) on its top face which is supported by
3.6 The Torso and Pelvis 105
2 2
10
6 5 9
4 4 11
7 3 3 8
3 5
5
1 1
(a) Front view (b) Rear view
1 Pelvis segment
17 18 2 Camera head
3 Arm
14 4 Shoulder joint flexion drive
20 5 Motor electronics
6 DC/DC converter
21 7 Inertial measurement system
8 Central control unit
9 Connector panel
10 Emergency stop pushbutton
11 Handle
12 Pelvis joint connecting flange
15 13 Main girder
16 14 Shoulder girder
15 Shoulder transverse tube
19 16 Shoulder joint flange
13 17 Head mount
18 Strut (head)
12 19 Strut (shoulder)
20 Fastening eye
21 Locating pin
(c) Trunk support structure
Figure 3.14: Mechanical design of the upper body
head strut (18) for increased tilting stiffness. In order to avoid damage of the robot
during experiments, it is secured by a safety rope. The rope is attached to the
robot by the fastening eyes (20). Lastly, two locating pins (21) are used to fixate
the upper body in the calibration jig (cf. Section 3.11.1).
The pelvis interfaces to the internal/external rotation joints of the legs and to
the upper body via the two-DoF pelvis joint. Using the pelvis joint, upper body
orientation is independent from leg motion except for flexion/extension. Pelvic
rotation can be utilized to increase the step lengths and provides an additional
means of compensation the total body angular momentum, as discussed in Sec-
tion 2.2.5. Leg swing induces high torsional moments into the pelvis structure.
The main body shown in Figure 3.14(a) must therefore show high stiffness to
(a) Pelvis base body
2
4 1
6
5
3
1 Pelvis segment
2 Upper body support structure
3 Adduction/abduction joint
4 Internal/external rotation joint
5 End stop
6 Pelvis joint connecting member
(b) Pelvis joint
Figure 3.15: Mechanical design of the pelvis segment and joint
ensure accurate foot placement with low distortion. The compact and lightweight
pelvis segment is manufactured from aluminum by investment casting10 . The
two-DoF pelvis joint is shown in Figure 3.15(b). The adduction/abduction (3)
and internal/external rotation joints (4) connect to the pelvis segment (1) and the
upper body support structure (2), respectively. The mechanical end stops (5) limit
the angular motion of the adduction/abduction axis and avoid collisions with
electronics components and cable overstretching.
3.7 The Arms

The arms of the robot have a rather minimalist design with three DoFs each.
Since manipulation tasks are not considered, they are used only for compensating
total body momentum to avoid slippage. Figure 3.16 shows the design of the
right arm. The housing of the shoulder flexion/extension drive (2) is flanged on
the support structure of the upper body (1). The shoulder drive (2) is a type ‘E’
servo actuator (Figure 3.9) with a single crossed roller bearing at its output. The
shoulder adduction/abduction joint (3) and the elbow flexion/extension joint (4)
10 See Section 3.10 for details on the design and manufacturing of the investment castings.
3.8 The Legs 107
3 2
5
8
1
6 7 1 Upper body support structure

2 Shoulder joint flexion/extension
3 Shoulder joint adduction/abduction
4 Elbow joint flexion/extension
4 5 Upper arm
6 Lower arm
7 Balance weight
8 Motor controller
Figure 3.16: Mechanical design of the arms
are driven by a type ‘F’ servo actuators with four-point contact ball bearings and
additional support bearings. The upper arm (5) and lower arm (6) are made of
thin-walled aluminum tubes. The connecting flanges to the servo actuators are
mounted with blind rivets and polyurethane adhesive bond. The lower arms (6)
are equipped with balance weights (7) at their ends. The upper arm holds the
motor controller (8) of the elbow joint.
3.8 The Legs

This section introduces the leg design which is considered the core component of
the mechanical structure. Section 3.8.1 introduces the hip joint, a spherical joint
composed of three serial drives. Sections 3.8.2 and 3.8.3 elaborate on the novel
drive mechanisms for the knee and ankle joint, respectively. Both designs reduce
the resultant leg inertia significantly, as discussed in Section 2.4. Section 3.12
concludes with a discussion of the mechanical design and compares different drive
mechanisms of the knee and ankle joints.
3.8.1 The Hip Joint

As stated in Section 2.6.1, the three-DoF hip joint is composed of serial drives,
forming a spherical joint with common intersection point. Particular care is
necessary in the conception and design of the hip joint. Besides sufficient actuator
performance, the static and dynamic stiffness of the connecting joint between
the leg and the pelvis is fundamental for the walking performance of the robot.
Figure 3.17 shows the mechanical design of the hip joint. The internal/external
rotation axis (2) is inclined 10° to the vertical axis to achieve a better power
distribution among the drives [46].
The hip joint drives are among the highest loaded joints of the robot. The
integration of these drives in a very narrow installation space is most challenging.
In addition, the hip joint structure and bearings must show high tilting rigid-
ity which is complicated by the large workspace of the flexion/extension axis.
Therefore, several design approaches were designed and compared in terms of
their static and dynamic properties. The chosen design shown in Figure 3.17a
comprises the internal/external rotation drive (2) with crossed roller bearing (type
‘B’, Figure 3.9) which is screwed to the pelvis segment (1). It is followed by the
adduction/abduction drive (3) and the flexion/extension drive (4) which are of
type ‘C’ and ‘A’, respectively. For high tilting rigidity and direct force flux, drives
(3) and (4) comprise four-point contact ball bearings at the output flanges and
additional support bearings (7). Both bearings (7) are thin-section deep groove
ball bearings in a non-locating arrangement. The fork-shaped rotation segment
(8) holds the adduction/abduction drive (3) and support bearing (7), surrounding
the hip joint adduction/abduction segment (9). Despite its slender shape, bending
and torsional stiffness of segment (8) are high enough.
Figure 3.17b explains the fixation of the hip joint in the pelvis segment. The
connecting flange of the internal/external rotation drive (2) is screwed on the
pelvis segment (1) together with the outer ring of crossed roller bearing (5). For
direct force transmission into the pelvis segment, motor housing (2) is additionally
supported by the polyurethane ring (6) inserted between the motor housing (2) and
bearing surface of pelvis segment (1). The viscoelastic properties of polyurethane
ring (6) enhance damping and provide an effective way to reduce structural
vibrations and oscillations of the swing foot.
3.8.2 The Knee Joint

As pointed out in Section 2.6.2, a linear drive acting as length-variable coupler
improves the mass distribution in the hip-thigh area, opposite to a conventional
design with Harmonic Drive gears. Hence thigh inertia is only marginally larger
compared to Johnnie but knee joint performance is increased significantly. In addi-
tion, using planetary roller screws instead of ballscrews leads to additional weight
savings due to the considerably higher load rating of these machine elements.
Figure 3.18a gives an overview of the one-DoF knee joint (1), located between
the thigh (2) and shank (4) segments. The servomechanism comprises the linear
actuator (5) and the absolute angular sensor (8) that enables the direct measure-
ment of the knee angle. The driving side of linear actuator (5) is attached to the
thigh segment (2) by the universal joint (6). The output side of actuator (5) is
connected to the shank (4) by the universal joint (7). Because of limited space
and moving parts in the knee joint area, all cables coming from the shank (4) are
guided by the protective cable conduit (9) to avoid cable jamming.
3.8 The Legs 109
6
2
b
8
3 1
4 2
9
bb
10
5
8
(a) Overview (b) Hip joint rotation axis
1 Pelvis segment 5 Crossed roller bearing 8 Hip joint rotation segment

2 Internal/external rotation drive 6 Polyurethane ring 9 Hip joint adduction segment
3 Adduction/abduction drive 7 Support bearing 10 Thigh segment
4 Flexion/extension drive
Figure 3.17: Mechanical design of the hip joint composed of three serial drives with
intersecting axes
Because of limited available space in the knee joint axis, the absolute angular
sensor (8) is shifted towards the thigh (2) using a parallel-crank mechanism. Fig-
ure 3.18b explains the sensor arrangement. The four-bar mechanism is composed
of the lever arm (11) mounted on the shaft of the sensor (8) and the coupler (12)
which connects the shank segment (4) and the lever arm (11) using spherical plain
bearings. The length of lever arm (11) equals the distance from the knee joint (1)
axis to the shank articulation point of coupler (12), while the length of coupler
(12) corresponds to the distance between the knee joint (1) axis and the shaft of
the absolute angular sensor (8). Figure 3.18b shows the two extreme positions
of the shank segment (4) for knee angles of -120° and 5°. The positions of lever
arm (13) and coupler (14) at a knee angle of -120° show that coupler (14) must be
cranked to avoid collisions with the angular gear (10) which is coaxially with the
knee joint.
Figure 3.18c shows the compact design of universal joint (7) which connects the
output link of knee joint actuator (5) to the shank segment (4). The spider (15)
is made from high-tension steel in order to keep its dimensions to a minimum.
Precision plain bearings (16) with adjusted clearance are used in the one axis
of the spider (15), where only minimal rotation due to elastic deformation of
the leg structure occurs. Since the main axis of rotation is subject to higher
rotational speeds, it is supported by the needle roller bearing (17) which can
1 Knee joint
2 Thigh segment
3 Hip joint
4 Shank segment
5 Knee joint linear actuator
6 Universal joint (thigh)
7 Universal joint (shank)
9 Protective cable conduit
10 Angular gear (ankle joint)
3 11 Lever arm (-120° position)
12 Coupler (-120° position)
5 13 Lever arm (5° position)
14 Coupler (5° position)
15 Spider
16 Plain bearing
17 Needle roller bearing
18 Axial bearing
19 Grease nipple
20 Shank segment axis
6 21 Knee joint bearing
22 Bearing flange (thigh)
23 Bearing flange (shank)
24 Locking ring & adjustment shims
25 Screws
5 2
2
8
13
8
14
7
11
9 4 12
1
10
4
1
(a) Overview (b) Absolute angular sensor
19
15
16
10 10
22 22
4 20 25 25
18 24 24
17 21 21
16 23 23
(c) Universal joint (shank segment) (d) Knee joint bearing structure
Figure 3.18: Mechanical design of the knee joint

3.8 The Legs 111
k
Output (driven) lin
Motor (driving) link
19 18
10 2 17
20 21 14
16
13 11
15 12
1
5 8 6
9
3 7
4
17 16
19 18
1 Stator 8 Locknut 15 Spider (thigh)

2 Motor housing 9 Roller screw bearing B 16 Plain bearing
3 Permanent magnet rotor 10 Incremental encoder 17 Steel raceway
4 Rotor receptacle 11 Roller screw shaft 18 Cantilever axis (thigh)
5 Cone clamping 12 Roller screw nut 19 Adjustment shim
6 Roller screw bearing A 13 Immersion tube 20 Protective cover
7 Bearing spacer 14 Universal joint (shank) 21 Cable grommet
Figure 3.19: Mechanical design of the knee joint actuator
accommodate high radial forces at very compact dimensions. It is complemented

by two axial bearings (18) which prevent the universal joint (7) from floating.
Each axial bearing (18) is composed of a bearing washer that is bonded onto the
shank segment (4), and a plastic thrust washer mounted on spider (15). The grease
nipple (19) simplifies maintenance of needle roller bearing (17). The universal
joint (7) is located in the shank segment (4) using the shank axis (20) which is
divided for assembly. It is inserted into the fitting bores from both sides of shank
segment (4) and fastened by a screw. Since the axis (20) also axially locates the
inner ring of bearing (17), its axial fixation in the shank is adjusted with shims
(not shown).
Finally, Figure 3.18d shows the bearing arrangement of the knee joint: it is
composed of two thin-section deep groove ball bearings (21) in an adjusted O-
arrangement. The outer ring of bearing (21) is press-fit into the bearing flange
(22) of the thigh segment. The inner ring is mounted on the bearing flange (23)
of the shank segment. Bearing flange (23) is designed as an integral part of the
angular gear (10), cf. Section 3.8.3. Bearing clearance is adjusted with shims (not
shown). The inner ring of bearing (21) is axially located by the locking ring (24)
and screws (25).
The mechanical design of the knee joint actuator is presented in Figure 3.19.
It is composed of the motor (driving) link which connects to the thigh segment
and the output (driven) link, connecting to the shank segment. In the driving
link, PMSM stator (1) is bonded into the motor housing (2) using a thermally
conductive epoxy. The motor housing (2) has cooling fins for natural convection.
The permanent magnet rotor (3) is bonded onto the motor receptacle (4). For
assembly convenience, the receptacle (4) is divided into two parts which are
clamped together using radially arranged set screws that engage into a V-groove.
The cone clamping (5) fastens the receptacle (4) on the roller screw shaft (11).
The rotating part of the driving link which consists of the roller screw shaft (11)
and motor receptacle (4), rotates on two angular contact ball bearings (6) and a
thin-section deep groove ball bearing (9) in a locating/non-locating arrangement.
Unlike the Harmonic Drive-based revolute joints, cf. Section 3.4.2, where the
locating bearing is placed next to the incremental encoder, locating bearing (6) of
the the roller screw shaft (11) resides on the opposite side of motor housing (2),
providing ideal force flux to the thigh segment. Locating bearing (6) is composed
of two matched single row angular contact ball bearings with internal pre-load.
This arrangement allows to accommodate high axial loads acting in both directions.
Bearings (6) are mounted immediately adjacent to each other and axially located
by bearing spacer (7) and locknut (8). The non-locating bearing (9) provides
additional radial support and ensures reliable working distance between the
magnetic ring and the scanning head of incremental encoder (10). Cover (20) and
cable grommet (21) protect the encoder (10) against ingress of solid foreign objects
and dust. The motor link is cardanically supported in the thigh segment by the
spider (15). Oil-free precision plain bearings (16) with minimal clearance are used
in both axes of the universal joint, as they are better suited for oscillating bearing
motions with small angles of oscillation than rolling bearings. They run on steel
raceways (17) which are bonded onto the bearing journals of the motor housing (2,
not shown) and spider (15). The spider (15) is located in the thigh segment by the
cantilever axes (18). To prevent the motor link from floating, axial clearance of
the bearings (16) is adjusted with shims (19) during assembly. The output (driven)
link consists of the roller screw nut (12), which is flanged on the immersion tube
(13). The figure shows the working principle of the drive. The roller screw shaft
(11) plunges into the immersion tube (13), changing the distance between the
upper and lower hinging points of the actuator. The output link connects to the
shank segment by universal joint (14), see Figure 3.18c.
3.8.3 The Ankle Joint

As shown in Section 2.6.3, the ankle joint axes differ clearly in their torque-
speed characteristics. Despite the higher development effort, the main reason
to implement a parallel mechanism is the significant reduction of the average
loads that are distributed almost equally among the motors. The low-inertia
design of this new drive mechanism places heavier component parts near the hip
joint. The mechanism is stiff, singularity-free throughout the working range, it is
back-drivable and shows only minimal backlash. The overall reduction ratio of
the mechanism ranges from 30 to 70. Due to the high complexity, the explanation
of the mechanism is split into two figures. Figure 3.20 gives an overview of the
entire drive mechanism and describes the transmission components on the thigh
segment, Figure 3.21 shows the spatial slider-crank mechanism and all other
transmission components located on the shank.
3.8 The Legs 113
4
5
6
7
2 1 Ankle joint 14 Motor shaft bearing A

2 Shank segment 15 Motor shaft bearing B
8 3 Foot segment 16 Support bearing
9 4 Thigh segment 17 Support bearing housing
5 Ankle joint motor 18 Pulley (driving)
6 Synchronous belt drive 19 Pulley (driven)
1 7 Angular gear 20 Cone clamping
8 Linear carriage 21 Incremental encoder
3 9 Coupler link 22 Protective cover
10 Stator 23 Cable grommet
11 Motor housing 24 Synchronous belt
12 Permanent magnet rotor 25 Idler (fixed)
13 Motor shaft 26 Idler (adjustable)
(a) Overview
11 21 5 4
18
10
25
23
24
22
26
19
2
16 20 18 17 15 13 12 14
(b) Ankle joint motor (c) Synchronous belt drive
Figure 3.20: Mechanical design of the ankle joint, part 1
Figure 3.20a shows the overall structure of the mechanism. The two-DoF
ankle joint (1) connects the shank (2) to the foot segment (3). It is actuated by
two spatial slider-crank mechanisms (8–9) with one DoF each. The motors (5)
are mounted on the thigh segment (4), close to the hip joint. Each slider-crank
mechanism is basically composed of the linear carriage (8) which translates along
the longitudinal axis of shank (2). The coupler link (9) connects carriage (8) and
the foot segment (3). Moving the linear carriages (8) on both sides of the shank (2)
in equal directions results in a flexion/extension movement of the ankle joint (1),
moving them in opposite direction yields an adduction/abduction movement. The
linear carriages (8) are driven by the motors (5) which connect to the slider-crank
mechanisms (8–9) via synchronous belt drives (6) and angular gears (7).
Figure 3.20b explains the design of the ankle joint motor. The structure is
almost identical to the Harmonic Drive-based revolute joints, cf. Section 3.4.2.
The PMSM stator (10) is bonded into the motor housing (11). The permanent
magnet rotor (12) is bonded onto motor shaft (13) which is supported by bearings
(14) and (15) in a locating/non-locating arrangement. Thin-section deep groove
ball bearings are employed for both locating (14) and non-locating bearing (15).
For better accommodation of the radial forces resulting from the synchronous
belt drive (6), the motor shaft (13) is additionally supported by the self-aligning
ball bearing (16). The ability of this arrangement to compensate dynamic angular
errors as well as axial misalignment avoids internal forces in the bearing system of
the motor shaft. Accordingly, a high-strength yet lightweight design of motor shaft
(13) is achieved by a two-part design: the inner part which holds the permanent
magnet rotor (12) is made from aluminum. Since the output part of the motor
shaft is subject to considerable rotating bending resulting from the radial forces
of the synchronous belt drive, it is made from high-tension steel. The driving
pulley (18) of synchronous belt drive (6) is fastened onto motor shaft (13) using
cone clamping (20). The support bearing housing (17) holds the bearing (16) and
provides protection against accidental contact of the pulley (18). The incremental
encoder (21) is mounted on the opposite side of the motor shaft (13) and protected
by the cover (22) and cable grommets (23).
Figure 3.20c shows the synchronous belt drive that connects the motor (5) to the
input shaft of the angular gear (not shown). The driving pulley (18) is fastened
onto the motor shaft (5). The driven pulley (19) resides on the input shaft of
the angular gear, coaxially with the knee joint. The pulleys are connected by the
synchronous belt (24). The fixed idler (25) increases the wrap angle of the driving
pulley (18). Accordingly, the adjustable idler (26) increases the wrap angle of the
driven pulley (19) and allows to set belt tension and compensate the tolerances of
synchronous belt (24).
Figure 3.21 illustrates all drivetrain components situated on the shank. Fig-
ure 3.21a shows the spatial slider-crank mechanism (8–9), located on both sides of
shank (2). The main components are the linear carriage (8), which moves in longi-
tudinal direction of shank (2), and coupler link (9) which connects carriage (8) to
the foot (3). The linear carriage (8) is guided by a pre-loaded linear recirculating
ball bearing and guideway assembly (32), which can accommodate loads from
all directions—except for the direction of motion—and moments about all axes.
The linear bearing unit comprises a single guideway (32) and two ball bearing
carriages which are screwed to carriage (8). Using two smaller instead of one
larger carriage reduces the mass and increases rating life. In addition, guideway
(32) is inclined by 20° to the sagittal plane to achieve a better loading condition.
The linear carriage (8) is driven by a planetary roller screw. The screw shaft (27)
and nut (28) are attached to the shank (2) and carriage (8), respectively. The
screw shaft (27) is supported by the axial angular contact ball bearing (29) and
needle roller bearing (30) in a locating/non-locating arrangement. Coupler link
(9) connects to the linear carriage (8) by a heavy-duty rod end (33) with reduced
clearance (5 − 10 µm) that can support high alternating loads and shock loads. The
3.8 The Legs 115
29
31 32 24
27 8 17
28 33 19 43
31 9 16
30 36
20
2 39
1 34 37
3 35
38
39
40
41 42 29 27
(a) Slider-crank mechanism (b) Angular gear
9
55
44
54
47
41
48
51
45 55
52 54
48
53
3 46 41
3 53
51
52
49
50
(c) Universal joint (coupler link) (d) Ankle joint
1 Ankle joint 30 Roller screw bearing B 43 Protective cable conduit

2 Shank segment 31 End stop 44 Universal joint adapter
3 Foot segment 32 Guideway 45 Spider
8 Linear carriage 33 Rod end 46 Cantilever axis
9 Coupler link 34 Universal joint (coupler link) 47 Set collar
16 Support bearing 35 Angular gear housing 48 Plain bearing
17 Support bearing housing 36 Angular gear input shaft 49 Flexion/extension axis
19 Pulley (driven) 37 Bevel gear assembly 50 Adduction/abduction axis
20 Cone clamping 38 Angular gear output shaft 51 Ankle joint bearing
24 Synchronous belt 39 Sealing ring 52 Locking plate
27 Roller screw shaft 40 Bellows coupling 53 Grease nipple
28 Roller screw nut 41 Locknut 54 Absolute angular sensor
29 Roller screw bearing A 42 Bearing spacer 55 Limit switch
Figure 3.21: Mechanical design of the ankle joint, part 2

other end of link (9) connects to the foot segment (3) by universal joint (34).
The slider-crank mechanisms (8–9) are arranged in the longitudinal direction of
the shank, perpendicular to the knee joint axis. It is therefore necessary to redirect
the axes of rotation of the motors (5) by angular gear units which are shown in
Figure 3.21b. Minimal backlash of these gears is crucial for servo performance
and accuracy, but can only be achieved by experience and very precise assembly.
Therefore, a commercial bevel gear (35–39) is employed in a customized version.
The precision gear unit is optimized for position servo applications and features
a lightweight aluminum housing (35) for good heat dissipation. The internal
gear design provides high efficiency ranging from 94 % to 97 % and minimal
angular backlash of less than 8 arc min [55]. The bevel gear unit is fitted with
hardened bevel gears (37) with spiral toothing, allowing for higher speed and
load rating and less vibration and noise. Moreover, the angular gears have amply
dimensioned ball bearings for a high degree of transmission accuracy. As the speed
range requires oil splash lubrication, the gear box is sealed with sealing rings (39).
Synchronous belt (24) connects the motor shaft (not shown) to the input shaft (36)
of the angular gear which is coaxially with the knee joint. The driven pulley (19)
of the synchronous belt drive is fastened onto the input shaft (36) using the cone
clamping (20). The self-aligning ball bearing (16) accommodates the radial forces
resulting from the synchronous belt drive and reduces the loading of input shaft
(36). Bearing (16) is mounted on the knee joint bearing flange using the support
bearing housing (17). The miniature bellows coupling (40) connects the output
shaft (38) of the angular gear to the roller screw shaft (27). It is characterized by
zero backlash, high torsional stiffness, low inertia and the ability to compensate for
axial, lateral and angular shaft misalignments. The coupling incorporates a blind
mate press-fit connection for assembly convenience. The roller screw shaft (27) is
axially located by the axial angular contact ball bearing (29), which is fastened
by the bearing spacer (42) and the locknut (41). The axial angular contact ball
bearing (29) is a double row unit with internal pre-load to accommodate radial
forces as well as high axial forces acting in both directions. It is optimized for
screw drives and characterized by high accuracy, high load carrying capacity, high
rigidity and low friction. The protective cable conduit (43) guides all cables in the
knee joint area and avoids cable jamming.
The large required workspace in the dorsiflexion/plantarflexion axis inhibits
the use of standard spherical plain bearings to connect the the coupler link (9)
and foot segment (3). Hence the universal joint (34) is developed as shown in
Figure 3.21c. The fork-shaped universal joint adapter (44) at the end of coupler
(9) clasps around the spider (45). The cranked shape of (44) is due to limited
space in the shank-foot area and avoids collisions with the foot (3) and shank.
The cantilever axis (46) is bolted into the foot segment (3) and holds the spider
(45). Both spider (45) and cantilever axis (46) are made from high-tension steel to
achieve extremely compact dimensions and the required mating material hardness
for the plain bearings (48). All axes of the universal joint run on oil-free bronze
bushings (48) which are assembled with minimal clearance (5 − 10 µm). Plain
bearings with solid lubrication are superior in abrasion resistance compared to
composite bushings. They can withstand heavy and shock loads and are suitable
3.9 The Feet 117
for low-speed oscillating motions at small angles of oscillation. Axial clearance of

the bearings running on cantilever axis (46) are adjusted with set collar (47).
Finally, Figure 3.21d concludes with the ankle joint that is designed as Cardan
joint with intersecting axes. The shank segment (2) holds the adduction/abduction
axis (50), the flexion/extension axis (49) is assigned to the foot segment (3). The
bearing arrangement is identical for both axes and comprises combined needle
roller/angular contact ball bearings (51) in an adjusted bearing arrangement, thus
the ankle joint is free from backlash. The bearings (51) are able to accommodate
radial and axial loads and allow to realize an adjusted bearing arrangement with
high radial load carrying capacity at limited radial space. The outer rings of the
bearings (51) rest against retaining rings, the inner rings are fixed with locking
plates (52) and locknuts (41), respectively. The latter are also used to adjust bearing
clearance. Grease nipples (53) at all bearings simplify bearing maintenance. The
absolute angular sensors (54) allow the direct measurement and control of the
joint angles, and are complemented by the limit switches (55). However, the
allowed angular range must be monitored by software, since the operating range
of the ankle joint is non-rectangular due to collisions between the shank (2) and
foot segments (3).
3.9 The Feet

As stated in Section 2.3, the main requirements on the foot structure include:
– High grip on different surfaces to ensure proper ground contact
– Passively compliant elements with good damping characteristics for shock
attenuation at initial contact and progressive spring characteristics at ground
contact to ensure effectiveness of stabilizing control
– Rigid mechanical structure to transmit the propulsive forces at a bandwidth
sufficient for system stabilization
– Compensate smaller unevenness
– Minimal mass
From these requirements, the design of the robotic foot shown in Figure 3.22a is
derived. The foot segment is connected to the shank (1) by the two-DoF ankle
joint (2) and actuated by the coupler links (3) via the universal joints (4). The foot
receptacle (5) holds the flexion/extension bearings of the ankle joint and interfaces
to the inner sensor flange of force/torque sensor (6), which is an integral, load-
bearing element within the forefoot. Details on the sensor design and function
are given in Section 4.2. The sensor (6) is shifted 40 mm towards the center of the
support polygon in order to achieve equal loading conditions in either direction.
Every corner of the sensor frame has a connecting flange (7), accepting the toe
joint actuator (8) at its front and the heel segment at its back. The toe joint (8)
divides the foot into forefoot and hallux.
Several mechanisms account for effective shock attenuation. Viscoelastic contact
elements (9) with good damping characteristic are used in analogy to the fat pads
in the human heel and toe for damping the impact transients and compensating
small unevenness. The design of the contact elements is described below in more
1 3
m 4
40 m
11
12
2 10
5 13
7
8
6
9
9
l
He e
foo t
9 Fore
ux
Hall
(a) Overview
α
R
l2
f ≈ t
24R
Tangent
14 Circular curve
f
lt /2
Transition curve
15 17 16 lt
(b) Contact element (c) Contact element geometry
1 Shank 7 Foot connecting flange 13 End stop

2 Ankle joint 8 Toe joint (actuator) 14 Contact element base body
3 Coupler link (ankle joint) 9 Contact element 15 Viscoelastic layer A
4 Universal joint 10 Heel bearing 16 Viscoelastic layer B
5 Foot receptacle 11 Hydraulic shock absorber 17 Outsole
6 Force/torque sensor 12 Four-bar linkage
Figure 3.22: Mechanical design of the feet

3.9 The Feet 119
detail. In addition, a passive heel segment is proposed for improved energy

dissipation during initial contact: similar to the human hindfoot, the passive
heel contributes to energy dissipation during initial contact by increasing the
time over which the foot is decelerated and energy is dissipated. For that reason,
it incorporates a hydraulic shock absorber (11) with adjustable damping. The
four-bar linkage (12) adjusts the strokes of heel and damper. The heel bearing (10)
consists of two spherical plain bearings in a locating/non-locating arrangement
and corresponds to the subtalar joint of the human foot. The mechanical end stop
(13) ensures proper transmission of the propulsive forces and stabilizing moments
while in ground contact. The feasibility of this new design has been proved in
preliminary tests with the robot Johnnie [164].
Figure 3.22c illustrates the multi-layered structure of these elements. The rigid
aluminum base body (14) interfaces to the heel and toe segments by a combined
parallel key and screw connection. Bonded onto contact element (14) are viscoelas-
tic layers (15,16) with different material properties, adjusted for better impact
attenuation. The viscoelastic layers (15) and (16) consist of Sylomer, a volume-
compressible PUR elastomer that is widely used in vibration isolation [45]. Layer
A (15) covers the major part of the contact area and is mainly relevant during mid
and terminal stance. An elastomer with comparatively low modulus of elasticity
is chosen due to the large contact area, where the loads are distributed among the
four contact elements. The material properties are adjusted experimentally to the
bandwidth of the ankle joint torque controller. Conversely, layer B is mainly active
during initial contact. The specific loads acting on this layer are significantly
higher due to the high landing impacts and the small contact area. Thus, an
elastomer with higher modulus of elasticity is used. The outsole (17) is bonded
onto the viscoelastic layer. It is made from wear-resistant, high-friction rubber
from orthopedic supplies.
The geometric shape of the heel contact elements permits a smooth rolling
motion of the feet. At initial contact the heel contacts ground and rolls on the
circular section of the contact element. If the straight section is tangent to the
circular section, the hallux touches down abruptly as the curvature at the tran-
sition point jumps from zero to the reciprocal of radius R. Therefore, in order
to prevent abrupt forces during the loading response phase, a transition curve is
fitted between the circular and straight section of the contact element. The transi-
tion curve is a clothoid, which is widely used to connect a straight and a circular
section of a rail track or highway. As illustrated in Figure 3.22c, its curvature
begins at zero at the straight section, increases linearly with its curve length and
ends at the curvature of the circular curve, 1/R. The vertical offset f between the
straight section and the circular curve can be calculated approximately from the
radius R and the transition length lt .
The functionality of the foot during the gait phases is illustrated in Figure 3.23.
At initial contact, Figure 3.23a, the heel segment protrudes from the foot sole.
Its passive motion increases the time over which the foot is decelerated and
during which energy is dissipated by the hydraulic damper. At loading response,
Figure 3.23b, the heel segment is retracted and more impact energy is dissipated
by deformation of the viscoelastic layers. The foot rolls from heel to toe and
(a) Initial contact (b) Loading response
(c) Mid stance (d) Terminal stance
Figure 3.23: Function of the heel and toe segments during one gait cycle. Besides
viscoelastic members, the passive heel segment with hydraulic damper contributes
to shock attenuation.
the hallux touches down. Ground contact during mid stance, Figure 3.23c, is
ensured by the heel and hallux. Then the robot’s weight is transferred to the
hallux and, finally, the heel lifts in terminal stance, Figure 3.23d. However, these
passive mechanisms of shock absorption are limited by the required bandwidth
for stabilizing the upper body orientation, see Section 2.3.
The modular mechanical configuration of the feet enables modifications of the
foot geometry (contact area size and shape, geometry of contact elements) by
simply replacing the contact elements. Moreover, it facilitates experiments with
different foot configurations. Figure 3.24a shows the default configuration of
Lola’s feet with active toe joints (8), using the lightweight rotary actuators type ‘G’,
cf. Figure 3.9. Figure 3.24b shows a configuration with passive toe joint. Because
the passive toe joint is more compact, the anthropometric position of the toe joint
can be fully kept. The toe joint bearings are designed in a locating/non-locating
arrangement. Due to the high radial loads, locating bearing (18) is a combined
needle roller/angular contact ball bearing which can accommodate axial loads in
both directions. Bearing (18) is fixed with locknut (19). A needle roller bearing
is used as non-locating bearing (20). A torsional spring (21) is employed that
supplies torque to restore the joint to its zero position. The absolute angular
sensor (22) acquires the toe joint angle, providing it for the trajectory planning
3.10 The Link Segments 121
18 19 22 21 20
8
(a) Active toe joint (b) Passive toe joint
23
8 Toe joint (actuator)

18 Toe joint bearing A
19 Locknut
20 Toe joint bearing B
21 Torsional spring
23 Hallux receptacle
(c) Monolithic foot
Figure 3.24: Modular structure of the feet, enabling experimental evaluation of differ-
ent configurations
algorithms. The simplest hallux configuration, the monolithic foot, is depicted in

Figure 3.24c. Here, the hallux contact elements are mounted on the receptacles
(23), providing a rigid connection to the forefoot.
3.10 The Link Segments

The component-related selection of construction materials and joining techniques
offers great potential for weight savings in many cases. Manufacturing major
structural components from fiber-reinforced polymer, e. g., carbon-fiber rein-
forced plastic (CFRP), is quite obvious because—if properly designed—CFRP
components have a higher specific strength and stiffness than Aluminum. On
the other hand, CFRP components are more elaborate in design and production.
Moreover, missing facilities for in-house manufacturing make CFRP components
rather costly. Instead, combining semi-finished CFRP tubes and attachment parts
made from aluminum are a more economical solution for simpler parts such as
the arm segments. But such multi-layered CFRP tubes seem to be available only
in larger customized batches. Magnesium, as the metal with the lowest density
(1, 8 g/cm3 ) of all construction metals, is excellently suited for use in lightweight
constructions. Despite good machinability, Magnesium and its alloys are mainly
cast materials due to limited cold forming properties. Complex-shaped single
copy Magnesium components can be manufactured by investment casting as de-
scribed below. But again, very high manufacturing costs are against Magnesium
castings. The robot Johnnie was made solely from various standard and high-
tensile Aluminum alloys. These materials have been proven successful during the
lab work, where accidental tipping over of the robot never caused serious damage
to components. Furthermore, subsequent modifications are possible and can be
conducted promptly in the in-house workshop. Consequently, using CFRP or
Magnesium alloys for the robot structure is abandoned but the tried and tested
aluminum alloys are used instead for most components. Less loaded parts and
covers are made from plastic (e. g., POM, ABS).
As stated in Section 3.2, lightweight design also affects the selection of joining
techniques, i. e., screwing, bolting, riveting and/or bonding. The weight analysis
of the robot Johnnie has revealed that screws make 4,7 % of the robot’s weight [46],
suggesting to reduce bolted flange connections to a minimum. In particular, other
joining techniques should be preferred for connections that need not to be loosened
for maintenance. Although bolted flange joints provide assembly convenience,
they contribute certain compliance to the robot structure. Thus, increasing the
stiffness of a flanged joint without increasing its mass is a worthwhile design
objective. More reliable transmission of forces and torques is achieved by EKagrip
shims [34]. These friction-enhancing shims provide microscale interlocking with
the joint surfaces, increasing the static friction coefficient by up to 300 %. EKagrip
shims are utilized in the highly loaded flange connections of the Harmonic Drive
Flexspline and the output flange, Section 3.4.2, as well as in the hip, knee and ankle
joint flanges (cf. Sections 3.8.1 to 3.8.3). On the other hand, joining aluminum
parts by bonding and/or riveting provides reliable force and/or material-locking
connections. Opposite to welding, the main benefits include the maintaining
of original material strength, no distortion and the possibility to connect non-
weldable, high-tensile Aluminum alloys or steel parts. As blind rivets are made
from Aluminum, they allow joining connections with virtually no increase in
weight. For adhesive connections of aluminum, polyurethanes are preferred over
epoxies due to their tough-elastic properties and good impact and peel resistance.
In addition, combined bonding and riveting connections have higher strength and
stiffness.
Generally speaking, prototype investment castings made from Aluminum are a
more economical solution than Magnesium castings. Nevertheless, this technology
enables the manufacture of very sophisticated, complex-shaped and thin-walled
parts with indentations, openings of any kind and complex rib structures. Since
there are almost no limitations in component shape, investment casting allows the
integration of various functions into a single part and the design of one-piece link
segments. It is possible to integrate attaching parts to reduce assembly complexity.
Thus, overlap due to bolted or riveted flange connections is avoided. Other than
“conventional” manufacturing techniques such as milling from the solid, prototype
investment casting allows for designing one-piece shell structures. Hence mainly
the part exterior is utilized to support the structural load [163], opposite to frame
3.10 The Link Segments 123
(a) Hip joint adduction segment (two parts, milling from the solid)
(b) Leg components (investment casting). Left to right: thigh base body, shank base body, ankle
joint connecting flange.
Figure 3.25: Examples of link segments, manufactured by milling from the solid or
investment casting
or truss structures. The main objectives of this construction technique, which is

also referred to as Monocoque design, are spreading the structural load evenly over
the part and reducing the overall workload, finally leading to additional weight
savings.
The prototype investment casting process starts with the manufacturing of the
investment casting model using the native 3D-CAD data of the finished part, con-
sidering allowances, shrinkage and mounting links for follow-up machining. The
models are manufactured by laser-sintering of plastic. The plastic master is then
embedded into molding material and molten out. After casting and cooling, the
casting blank part is cleaned and ready for after-treatment and inspection. This
includes heat treatment to adjust the mechanical characteristics, blank measure-
ment to check distortion and shrinkage, and X-ray imaging to check for internal
defects. The final CNC machining of the casting blank is essential as the castings
only become fully usable by the machining of all functional surfaces.
Figure 3.25 shows examples of link segments, manufactured either by milling
from the solid or prototype investment casting. Figure 3.25a depicts the fork-
shaped hip joint adduction/abduction segment which surrounds the hip joint
flexion/extension drive, cf. Figure 3.17. It is composed of two parts which are
manufactured by milling from the solid and joined using dowel pins and screws.
The left side of the assembly interfaces to the adduction/abduction drive by a
parallel key and screws. Figure 3.25b shows three components manufactured by
investment casting with machined functional surfaces: the main bodies of the
thigh and shank, and the ankle joint connecting flange. The castings incorporate
several functions: for example, the thigh base body has connecting flanges to the
hip and knee joints, bearing seats for the knee joint actuator, the flange for the
knee joint angular sensor, and two mounting surfaces for motor controllers with
cooling fins on their back. The shank base body comprises the connecting flange
to the knee joint, bearing seats for the ankle joint, and all mounting surfaces and
bearing seats for the ankle joint drive mechanism. The one-piece design allows to
combine all corresponding surfaces in a single component, which are machined
in a single setting in order to keep the geometric tolerances. This design and
manufacturing approach further ensures accuracy of kinematic parameters (e. g.,
location and orientation of neighboring links) which is necessary for precise and
robust trajectory planning and balance control.
3.11 Parameter Identification

This section deals with the identification of the kinematic and inertial parameters
of the robot after assembly. Determining the angular offsets of the link position
sensors is the main objective of kinematic calibration procedure described in Sec-
tion 3.11.1. The identification of the rigid-body inertial parameters, Section 3.11.2,
aims at obtaining an accurate model for multibody simulation and stabilizing
control.
3.11.1 Kinematic Calibration

The kinematic parameters of the robot include geometric parameters, such as the
location of the local coordinate systems of neighboring links, the location of the
stereo cameras relative to the upper body coordinate system, as well as angular
offsets of the link position sensors. For the robot geometry, it can be assumed
that all significant dimensions are within the respective tolerances for linear
and angular dimensions: particularly, the one-piece design of the leg segments
and the machining of all corresponding surfaces in a single setting ensures high
geometric accuracy and tight tolerances within few tenths of a millimeter, see
Section 3.10. On the other hand, the accurate identification of the angular offsets
of the absolute link position sensors is fundamental to the control performance.
The absolute angular sensors are mounted in arbitrary angular positions, thus
3.11 Parameter Identification 125
Posture 2
Posture 1
Posture 3
Robot
Calibration jig
Figure 3.26: Schematic diagram of the calibration procedure for closed-loop calibra-
tion of the joint angles
calibration for offset determination is required. The procedures of kinematic

calibration of robotic manipulators can be categorized in two general methods
[71]:
Open-loop calibration relies on external metrology to measure the posture of a
link chain, for example a camera-based motion capture system or a high-
precision inertial measurement unit mounted on the end link.
Closed-loop calibration physically constrains the end link in all degrees of free-
dom in predefined positions and orientations, substituting actual measure-
ments of the end link postures.
Due to lack of a motion capture system, a closed-loop calibration procedure is
chosen, using a calibration jig11 for accurate closed-loop calibration of the joint-
based parameters.
Figure 3.26 illustrates the method to identify the angular offsets of the encoders.
The calibration jig fixes the pelvis segment as the root link of the cantilever serial
leg structure. The upper body is positioned using the locating pins on the upper
body support structure (item (21) in Figure 3.14c). The feet can be mounted in
three different positions, utilizing the locating pins on the foot receptacle (item (5)
in Figure 3.22a). Hence the calibration jig enables the joint angle measurement in
three different leg postures to increase the calibration accuracy. The toe joint is
calibrated by another jig. Since it is not considered in the leg posture, the joint
angles of the remaining six-DoF serial link chain can be calculated by inverse
kinematics from the known position and orientation of the forefoot.
3.11.2 Link Inertial Parameter Estimation

The rigid-body inertial parameters are necessary for the multibody system used
for off-line simulation of the robot, and online trajectory planning. Moreover, the
inertial parameters enter the derivation of simplified models used for stabilizing
the robot. Thus, accurate knowledge of the inertial parameters is crucial for
walking performance and stability.
While the joint angles can be calibrated as described in the previous section,
the experimental identification of the inertial parameters is more involved. For
11 The calibration jig was designed by Markus Schwienbacher
Table 3.2: Mass distribution of the robot and percentage of the overall weight. The
amount of correction mass per segment was determined by weighing the robot. The
total masses per segment are the basis for the CAD-based approach for estimating
the full inertial parameters.
Correction mass [kg] Total mass [kg] Percentage [%]
Segment Left Right Left Right Left Right
Arms
Shoulder 0.0 0.0 0.561 0.561 0.92 0.92
Upper arm 0.205 0.198 1.900 1.891 3.12 3.11
Lower arm 0.0 0.0 1.486 1.486 2.44 2.44
Subtotal 0.205 0.198 3.947 3.938 6.48 6.47
Legs
Hip rotation 0.0 0.0 1.844 1.844 3.03 3.03
Hip adduction 0.0 0.0 2.838 2.838 4.66 4.66
Thigh 0.309 0.393 6.419 6.503 10.54 10.68
Shank 0.125 0.125 3.403 3.403 5.59 5.59
Ankle 0.0 0.0 0.138 0.138 0.23 0.23
Forefoot 0.027 0.032 1.643 1.648 2.70 2.71
Toe (active) 0.013 0.013 0.797 0.797 1.31 1.31
Subtotal 0.473 0.563 17.082 17.171 28.06 28.20
Upper body
Head tilt 0.0 1.013 1.66
Head pan 0.0 0.249 0.41
Torso 0.597 9.884 16.23
Subtotal 0.597 11.146 18.31
Pelvis
Pelvis rotation 0.166 1.316 2.16
Pelvis adduction 0.411 6.281 10.32
Subtotal 0.577 7.597 12.48
Full body total 2.612 60.881 100.0
example, using the feet as proof masses, the masses, centers of mass and inertia
tensors relative to the link coordinate systems can be estimated by measuring the
joint torques, velocities and accelerations for given leg trajectories. Because of
limited mobility and the lack of accurate joint torque sensing, it is impossible to
estimate all inertial data [71]. Although the incomplete data may be sufficient
for fixed-base manipulators, all inertial parameters are required to calculate
whole-body dynamics of a humanoid robot with unilateral foot-ground contact.
Therefore, a CAD-based approach is proposed for estimating the full inertial
parameters. Accurate modeling of all purchased components is accomplished by
weighing the parts and adjusting the material properties of the CAD model. Still,
the calculated mass of the link segments can only be a rough estimate as long
as cable harnesses are not considered. Therefore, all components, including the
cables, are weighed during assembly. While the mass of components manufactured
3.12 Discussion 127
by milling and/or turning can be determined reliably from the native CAD data,
the castings are weighed and the material density is adjusted in the CAD model.
A large portion of the robot’s mass is due to the joint servo actuators. To take
account for an accurate mass distribution, all drives are weighed after assembly.
For the output side of the drives, the numerically obtained mass from CAD data is
sufficient. The link masses are adjusted in the CAD model by adding solids with
adjusted material density, thus, the inertial parameters can be assumed to be close
to reality. Finally, the new inertial parameters are determined for the multibody
simulation model of the robot (Section 3.3.1).
Table 3.2 shows the mass of each segment, the percentage of the total mass and
the amount of correction masses. The total amount of correction masses is 2.612 kg,
corresponding to an initial modeling error of 4.3 %. The largest correction masses
are applied to the thighs, pelvis and upper body, where most of the electronics
and cable harnesses are situated.
3.12 Discussion
This section discusses some of the most relevant design features of the mechanical
structure. Section 3.12.1 treats the mass distribution of the robot. The chosen
designs of some leg joints are compared against other approaches in Section 3.12.2
regarding their effect on the dynamic performance of the overall system.
3.12.1 Mass Distribution

The total mass of the robot is 60.88 kg. The contribution of the individual links
given in Table 3.2 resembles the typical mass distribution of a biped walking robot.
Contrary to humans, where most of the mass resides in the upper part of the body,
the legs make the largest portion of the robot’s mass. Weighing 17.082 kg and
17.171 kg, the legs account for approximately 56 % of the total mass. Note that
the portion of the highly accelerated segments (thigh, shank, ankle, forefoot, toe)
is approximately 12.45 kg per leg, or 20.4 % of the total mass.
3.12.2 Analysis of Dynamic Performance

Consistent lightweight design of the mechanical structure with high effective stiff-
ness and natural frequencies obviously contributes to better dynamic properties
of the robot. As discussed in Section 2.4, greater acceleration capability of the
locomotor system is further achieved by systematically minimizing the inertia of
the individual links, as well as the resultant inertia of the whole kinematic chain.
These measures go along with lower nonlinear effects due to varying inertia, which
is expected to support the performance and robustness of balance control.
In order to asses the effectiveness of this approach, meaningful criteria are
necessary. Numerous dynamic performance criteria have been established for
fixed-base robotic manipulators [6], but few address legged robots. For example,
Bowling [15] proposes a measure of how well a legged robot uses the contact
forces to accelerate itself under consideration of the kinematic constraints asso-

ciated with unilateral ground contact. All these indices are suggesting that the
acceleration capability of the robot’s locomotor system is, to a large extent, a
function of the resultant leg moment of inertia.
Static and dynamic stiffness of the hip joint assembly is analyzed with special
consideration of the gear stiffness. Furthermore, the contribution of the imple-
mented drive mechanisms of the knee, ankle and toe joints to the resultant leg
moment of inertia is compared to other designs that have been realized in actual
robots or proposed in the literature. For a more reliable comparison, only compo-
nents of the drive mechanisms are considered. The effect of varying inertia due to
knee motion is analyzed for the ankle and toe joints. The mass and inertia data are
obtained numerically from CAD data throughout the analysis. Note that actual
mass and inertia data are only available for the implemented designs. Due to lack
of available reference data, they are estimated for all other design variants. Since
the estimation is done on basis of the available CAD data and experience, the data
can nevertheless assumed to be realistic.
Hip Joint Stiffness
Static and dynamic stiffness of the hip joint considerably affect accuracy of foot
placement. Even small (static) forces applied to the feet yield high torque loadings
of the hip joint due to the long lever arm of the foot contact points (≈ 1, 000 mm).
The torsional stiffnesses of the gear units in the adduction/abduction (Harmonic
Drive HFUC 25-100) and flexion/extension axes (Harmonic Drive CSG 32-50) are
k1 = 3.1 · 104 Nm/rad and k1 = 5.4 · 104 Nm/rad, respectively [59]. For example, a
load torque of 100 Nm causes a torsional distortion of the Flexspline of 3.226 ·
10−3 rad and 1.852 · 10−3 rad in the adduction/abduction and flexion/extension
axes, respectively, resulting in a foot displacement of 3.226 mm in the lateral,
and 1.852 mm in the fore-aft direction. For comparison, Johnnie incorporates
Harmonic Drive HFUC 25-80 gears in the flexion/extension joints, resulting in a
foot displacement of 3.226 mm in the fore-aft direction under the same loading
conditions.
As a result of the considerations in Section 2.4 and Footnote 6, lower link iner-
tias advance the natural frequency and antiresonance of the drive mechanisms.
In the following, the antiresonance frequency of the adduction/abduction and
flexion/extension drives are analyzed using the resultant inertia of the stretched
leg as a worst case assumption. For the adduction/abduction the resultant inertia
amounts to 3.62 kgm2 , and 3.631 kgm2 for the flexion/extension axis. Using the
torsional stiffnesses of the gear units in Eq. (3.17) on page 89, the antiresonances
of the adduction/abduction and flexion/extension drives are fn,add = 14.73 Hz and
fn,fl = 19.41 Hz, respectively. Thus, antiresonance of the adduction/abduction gear
is marginally smaller than the manufacturer’s recommendation, see Footnote 6,
but the antiresonance of the flexion/extension gear is well above this value. In con-
clusion, it was the right decision to choose the larger gear for the flexion/extension
drive. The high torsional stiffness significantly contributes to the accuracy of
foot placement even under high loading. Note, however, that these calculations
only consider the gear stiffness. The actual foot displacements are affected by the
stiffnesses of the link structure and its bearings and thus considerably higher.
3.12 Discussion 129
Table 3.3: Evaluation of different

knee joint mechanisms
Design Thigh inertia∗ [kg m2 ] Mass† [kg]
K1.1 0.585 3.009

K1.2 0.360 3.450
K2 0.150 3.560
K4 0.085 1.931
∗ For better comparison, only the drive mecha-
nisms are considered, any structural compo-
nents are neglected.
† Estimation (except design K4)
Knee Joint Mechanism

Table 3.3 compares the mass and thigh moment of inertia of four knee joint
mechanisms. Design variant K1.1 corresponds to solution K1, Figure 2.12a. The
mechanism consists of a Harmonic Drive-based servo actuator (type ‘A’, Figure 3.9)
that is mounted coaxially with the knee joint axis. This design obviously has the
largest moment of inertia because both motor and gear unit are located in the knee
joint. Design variant K1.2 is similar to the previous design but the motor is shifted
close to the hip joint. The Harmonic Drive-based gear unit is coaxially with the
knee joint axis and connected to the motor by a synchronous belt drive. The thigh
inertia decreases by 38.5 % but additional components (auxiliary transmissions,
shafts, bearings, housings) increases the mass of the mechanism by 14.7 %. Design
variant K2 corresponds to a patent application by Honda [51] and is shown
schematically in Figure 2.12b. It employs the same Harmonic Drive-based servo
actuator as variant K1.1 but the entire actuator is located close to the hip joint.
The knee joint is actuated by a coupler link, forming a parallel crank mechanism.
The thigh inertia reduces significantly by 74.4 % compared to design K1.1. On
the other hand, the estimated mass increases because of additional transmission
components such as heavy-duty bearings of the gear unit output flange and
coupler link. Finally, design variant K4 represents the implemented mechanism,
cf. Sections 2.6.2 and 3.8.2. The four-bar mechanism with length-variable linear
drive complies Figure 2.12d. The thigh inertia and the mass are reduced by 85.5 %
and 35.8 %, respectively.
In summary, the chosen knee joint mechanism K4 has shown advantageous
for the dynamic performance of the leg apparatus because it provides by far the
largest reduction in inertia and mass compared to well-established designs, such
as K1.1 and K1.2.
Ankle Joint Mechanism
Table 3.3 compares the mass and posture-dependent shank inertia of different
mechanisms for ankle joint actuation. The ankle joint in design variant A0 is
actuated in a serial arrangement. The considered drive mechanism is characterized
by reduced shank inertia compared to similar designs (cf. Section 2.6.3) and
covered by two patents by Honda [50, 119]. Harmonic Drive-based gear units
are mounted coaxially with the flexion/extension and adduction/abduction axes
of the ankle joint. The motors are located near the knee joint and connect to
Table 3.4: Evaluation of different ankle joint mechanisms

Shank inertia w.r.t. hip joint∗ ,
Design Thigh inertia∗ dependent on knee angle [kg m2 ] Mass†
[kg m2 ] 0° 30° 60° 90° [kg]
A0 — 1.743 1.635 1.341 0.939 3.12

A2 — 1.032 0.930 0.740 0.510 2.42
A3.1 — 0.932 0.898 0.770 0.592 2.44
A3.2 0.058 0.654 0.614 0.520 0.408 3.14
∗ For better comparison, only the drive mechanisms are considered, any struc-
tural components are neglected.
† Estimation (except design A3.2)
the gear units via synchronous belt drives. An additional bevel gear redirects
the motor axis of the adduction/abduction axis. Despite its simple kinematic
structure, this drive mechanism is unfavorable with regard to leg dynamics. The
long transmission distance between the motors and gear units leads to a rather
high mass and shank inertia is very high due to the large perpendicular distance of
the gear units to the hip joint. The mass distribution also results in a high posture
dependence. At a knee angle of 90° the shank inertia decreases by 23 % compared
to the stretched leg configuration. Design variant A2 corresponds to Figure 2.15b,
a detailed description is given in [113]. The ankle joint is actuated by a parallel
mechanism composed of two linear actuators acting as length-variable couplers.
The actuators are connected to the shank and feet by universal joints. Design A2
has the lowest mass of all considered mechanisms because its structure is very
simple and requires only a small number of components. Compared to design
variant A0, the shank inertia reduces to 59.2 % in the stretched leg configuration.
The variation of shank inertia over the knee joint workspace is 22.3 %.
The structure of design variant A3.1 is shown in Figure 2.16a. Two spatial
slider-crank mechanisms mounted on the shank drive the ankle joint in a parallel
arrangement. The motors are placed close to the knee joint and drive the linear
carriage through synchronous belt drives. Hence the entire drivetrain is mounted
on the shank, similar to the Johnnie design [47]. The mass of the mechanism is
comparable to design variant A2. The shank inertia is 53.5 % in the stretched
leg configuration compared to design variant A0, and varies by 19.1 % over the
relevant knee angles. Finally, design variant A3.2 is the mechanism that is adopted
to the ankle joint of Lola, cf. Section 3.8.3 and Figure 3.20. The kinematic structure
(Figure 2.16b) is identical to design A3.1 but the motors are placed on the thigh,
close to the hip joint. Power transmission from the motors to the planetary roller
screws occurs by synchronous belt drives and angular gears that are coaxial with
the knee joint. The higher mass compared to design A3.1 results from the longer
transmission distance between the motors and screws. Although the mass is
similar, the contribution to the resultant leg inertia reduces to 40.8 % of the serial
arrangement A0 (stretched leg configuration). The shank inertia varies by only
17.1 % over the knee angle.
The analysis shows that drive mechanism A3.2 drastically improves the ac-
celeration capability if the legs. Although the mass is comparable to the serial
Table 3.5: Evaluation of different toe joint configurations

Foot inertia w.r.t. hip joint∗ ,
Design dependent on knee angle [kg m2 ] Mass∗
0° 30° 60° 90° [kg]
monolithic foot 0.367 0.369 0.323 0.242 0.389

passive toe joint 0.503 0.505 0.441 0.328 0.537
active toe joint 0.830 0.833 0.727 0.542 0.912
∗ For better comparison, only the hallux is considered, the remaining foot
structure is neglected.
arrangement, which is found in most humanoid robots, it has by far the lowest
moment of inertia with respect to the hip joint. Also, the effects of varying shank
inertia are lower than in the other designs.
Toe Joint Mechanism
Table 3.5 compares the contribution of different foot configurations to the resultant
leg inertia. The mechanical design of the monolithic foot structure, passive and
active toe joint are shown in Figure 3.24.
Unsurprisingly, the monolithic foot structure, Figure 3.24c, has the lowest
inertia and mass. The inertia variation over the knee workspace is 34 %. The mass
of the passive toe joint illustrated in Figure 3.24b is 138 % of the monolithic foot,
likewise, the inertia increases to 137 % in the stretched leg configuration. With
34.8 %, the inertia variation is comparable. For the active toe joint, Figure 3.24a,
the mass and resultant inertia for the stretched leg configuration increase to 234 %
and 226 % of the monolithic foot, respectively. The inertia variation accounts for
34.7 % over the knee workspace.
Although a comparison between the different foot concepts is quite limited,
the analysis shows that the effort for improving the acceleration capability of
the locomotor system is little counteracted by the active toe joint. However, the
advantage of more flexible and natural gait generation and the potential to unload
other leg joints (see Section 2.2.6) outweigh the disadvantages.
3.13 Chapter Summary

This chapter detailed on the mechanical design of the robot. Distinguishing char-
acteristics are the redundant kinematic structure with seven-DoF legs, consistent
lightweight construction and a modular joint design with high power density.
Furthermore, mass distribution of the legs is improved to achieve good dynamic
performance:
Design calculations: fundamental design data are obtained from a multibody
simulation model and form a basis for dimensioning the mechanical struc-
ture and servomechanisms. Torque-based selection of the servomechanisms
and their components aims at meeting the static and dynamic load require-
ments. It is complemented by the elastodynamic analysis of the drive mech-
anisms to ensure high dynamic performance of the locomotor system. Most
structural components are designed using an empirical design approach.
Topology optimization is applied to components subject to complex loading.

Since structural vibrations can easily destabilize the robot, dynamic stiffness
of the leg links is analyzed by numerical modal analysis.
Modular joint design: the constraints of human dimensions and mass demand
for the development of very compact and lightweight servo actuators. The
rotary drives integrate permanent magnet synchronous motor, Harmonic
Drive gear, sensors and other components in very compact units with high
power density. For minimal mass and better mass distribution, the main
structure of the robot is non-modular. But the drives are built on the unit
construction principle, simplifying design and manufacturing.
High-dynamic leg structure: special emphasis was put on an improved mass
distribution of the leg apparatus to achieve good dynamic performance. Mass
distribution in the hip-thigh area is improved significantly by employing
a roller screw-based linear actuator in the knee joint. The ankle joint is
actuated by a parallel mechanism of two linear actuators whose motors are
mounted on the thigh. Thus, a large part of the actuator mass is be shifted
close to the hip joint rotational axis, resulting in high-dynamic behavior of
the legs. In addition, the base bodies of pelvis, thigh, shank and foot are
designed as investment castings. Prototype investment casting allows for
one-piece shell structures that result in very sophisticated, complex-shaped
and lightweight parts with high effective stiffness.
The feet: the toe joint divides the foot into forefoot and hallux. Despite the
lightweight design, the feet have a rigid mechanical structure to reliably
transmit the propulsive forces for system stabilization. The modular me-
chanical configuration allows replacement of the heel and toe segments
and/or other components. Several mechanisms are implemented for impact
attenuation: viscoelastic contact elements with good damping characteristic
are used in analogy to the human fat pads. Similar to the human hindfoot, a
passive heel segment with hydraulic shock absorber contributes to energy
dissipation during initial contact. Proper ground contact is ensured by an
outsole made from wear-resistant, high-friction rubber.
Parameter identification: a calibration jig is used to determine angular offsets of
the link position sensors in a closed-loop calibration procedure. In order
to realistically represent the mass distribution in the multibody simulation
model, all components including cable harnesses are weighed during assem-
bly. The link masses are adjusted in the CAD model and corrected inertial
parameters are adopted into the robot simulation.
Evaluation of dynamic performance: compared to other design variants, the
drive mechanisms of the knee and ankle joints by far have the lowest contri-
bution to resultant leg moment of inertia. Thus, acceleration capability of
the legs is improved significantly. Although the active toe joints little coun-
teract the low-inertia design of the legs, these disadvantages are outweighed
by the ability to generate more versatile gait patterns and the potential to
unload other leg joints.
4 Sensor System
The machine has no feelings, it feels no fear and no
hope [. . . ] it operates according to the pure logic of
probability. For this reason I assert that the robot
perceives more accurately than man.
—Max Frisch, Homo Faber
This chapter introduces the sensor system of the robot. The joint sensors for
motor control and link position sensing are explained in Section 4.1. The custom-
developed force/torque sensor for measuring the ground reactions is presented
in Section 4.2. Finally, the motivation for employing a high-precision inertial
measurement system is given in Section 4.4.
4.1 Joint Sensors

Each joint is equipped with an incremental rotary encoder, an absolute angular
sensor for direct link position sensing and a limit switch. For the rotary servo
actuators, all sensors are placed inside the housing, assembled as space-saving as
possible. The arrangement is exemplified in Figure 3.10.
Typical accuracy requirements on the encoder technology used are sufficient for
this application. However, the range of suitable sensors is severely limited by the
tight packaging goals and weight restrictions. Standard housed shaft-end encoders
come with integral bearing, hence the measuring accuracy is insusceptible to
alignment errors. Although easy to install and robust against dirt and vibrations,
the industrially ruggedized design makes such encoders too heavy and bulky.
Instead, kit encoders without integral bearing and housing must be used. In
this case the precision of the encoder mechanical assembly (concentricity of
the disk and scanning head, etc.) is crucial to measurement accuracy. Hence
encoder installation and alignment are very elaborate and require sophisticated
measurement equipment.
The incremental rotary encoder on the motor shaft utilizes the giant magnetore-
sistive effect to measure rotation by detecting the magnetic field and its changes.
Figure 4.1a shows the scanning head and scale graduation of this sensor. Magne-
toresistive encoders are characterized by a high resolution, robustness, resistance
against humidity, dust and oil, and a high-dynamic response. The scanning head
is a very compact PCB1 module with integrated signal conditioning electronics,
enabling simple assembly and adjustment without further calibration. The scale
graduation of an incremental magnetic encoder is a multi-pole magnetic ring.
In addition to software-based monitoring, each joint incorporates a limit switch
to avoid problems like self-collisions and cable wind-up. The sensor shown in
1 Printed circuit board
133
134 4 Sensor System
(a) Incremental rotary encoder (b) Limit switch with cam
(c) Absolute angular sensor, type 1 (d) Absolute angular sensor, type 2
Figure 4.1: Joint sensors
Figure 4.1b is composed of a photoelectric fork sensor and a cam that is connected
to the gear output. The allowed angular range is determined by the cam which is
different for each degree of freedom. When a limit is encountered motor voltage is
interrupted inside the motor controller module with minimum software overhead
before reaching the joint’s mechanical end stops.
Figures 4.1c and 4.1d show the angular sensors for load-side feedback. Both
sensor determine the absolute joint angle within a single revolution. The rotary
encoders of type ECI are manufactured by Heidenhain and typically used as
resolver replacement in industrial motor control. The sensors of type 1 and 2 differ
in the size and resolution. Type 1 shown in Figure 4.1c has a diameter of 37 mm
and a resolution of 16 bit per revolution and is used in the ankle and toe joints.
The other joints employ sensor 2, Figure 4.1d, which has a diameter of 65 mm and
a resolution of 17 bit. Both sensors have a measurement accuracy of 0.1°, a cutoff
frequency 6 kHz and a latency for continuous sampling of position values of 5 µs.
Hence the delay in position feedback is minimal and velocity measurement can
be done by numerical differentiation. Position values and configuration data are
transmitted via the digital, bidirectional EnDat interface. ECI encoders utilize
the inductive measuring principle and come without integrated bearing. The
4.2 Force/Torque Sensors 135
(a) Ground reaction forces (b) Ground reaction moments

400 140
120 mx
200 my
100 mz
0 80
60
Moment [Nm]
-200
40
Force [N]
-400 20
0
-600
-20
-800 fx -40
fy
-60
-1000 fz
-80
-1200 -100
11 12 13 14 15 16 17 11 12 13 14 15 16 17
Time [s] Time [s]
(c) FFT ground reaction forces (d) FFT ground reaction moments
200 25
mx
fx
my
fy
fz 20 mz
150
Amplitude [–]
Amplitude [–]
15
100
10
50
5
0 0
0 10 20 30 40 50 0 10 20 30 40 50
Frequency [Hz] Frequency [Hz]
Figure 4.2: Above, ground reaction forces and moments for a walking speed of
5.55 km/h obtained from simulation; below, frequency spectrum of the contact
wrench.
encoder shaft is fixed to the output flange using an elastomer coupling, as shown
in Figure 3.10, and the encoder flange is fixed to the motor housing. Besides
rotational speed, the attainable accuracy depends on the air gap between rotor
and stator and the mounting conditions. The air gap between the scanning PCBs
is adjusted with shims during assembly. As changes in the air gap, e. g., due to
thermal expansion of the sensor shaft, or wobbling of the rotor PCB, affect signal
quality. These errors are minimized by the short distance between encoder shaft
bearing and the rotor PCB.
4.2 Force/Torque Sensors

This section presents the design of the six-axis force/torque sensor based on the
fundamental design concept developed in Section 2.7.3. Figure 4.2 shows the
ground reaction forces and moments for walking at a speed of 5.55 km/h obtained
136 4 Sensor System
εxx ε̄SG
εxx
B
dA
x ε̄SG
Strain-sensitive area F
y
x
x
−ε̄SG
Figure 4.3: Cantilever beam as basic transducer geometry. Left, high sensitivity is
achieved by applying the strain gauges close to the clamping; right, the strain gauge
averages strain over the area covered by the grid.
from the multibody simulation model by Buschmann, cf. Section 3.3.1. Using
these data, the measurement range of the force/torque sensor is derived:
 
 500  100
 
F = ±  100  [N] ; M = ± 120 [Nm] . (4.1)
   
1200 50
   
Uchiyama et al. [189] propose a systematic procedure for the structural design
of a robotic force/torque sensor: the condition number of the strain-compliance
matrix serves as a measure for the maximum possible error in the measured forces
and moments. It is an indicator for the isotropy of structural compliance of the
transducer body with respect to sensor strain, and is equal to one in the case of
perfect isotropy. But the sensor performance is also affected by the relationship
of beam length l and sensor size L = r + l (cf. Figure 4.4) and other factors such
as stiffness, eigenfrequencies, decoupling, linearity, and hysteresis [52]. Lastly,
in order to obtain a wide enough bandwidth, the natural frequencies of the
transducer body should be higher than the frequencies of motion. To this end,
a fast Fourier transform (FFT) is conducted on the ground reaction forces to
provide an estimation of the frequency range of a typical gait pattern. The FFT
of the ground reaction forces and moments are shown in Figure 4.2(c) and (d),
respectively. These results indicate that the frequency spectrum of the reference
walking pattern ranges up to 25 Hz.
The following sections describe the mechanical design, sensor electronics and
the calibration and verification procedure. Most of these results are lifted from
the Diploma thesis by Schwienbacher [167].
4.2.1 Optimized Shear Beam Geometry for High Sensitivity

The most fundamental force transducer shown on the left of Figure 4.3 is the
cantilevered beam with strain gauges applied to its surface. The force F induces
surface strain ε that increases arithmetically towards the clamping. Hence the
fz
2
bB , E, Iyy x
fz bM
lM
z
bB
bB bM l
l lM
E, Iyy hM x
y
Q
E, IM
lM
m
f l
m0 = − 2z
u ⊕
r l x
z
u ⊕
Figure 4.4: Left, mechanical model of a shear beam (index B) supported by a mem-
brane (index M); right, reduced planar model of the shear beam. The elastic rela-
tions (shear Q, moment m, slope xu , deflection u) are shown qualitatively.
strain gauges are placed close to the clamping to achieve high sensitivity. The
transducer sensitivity βT obviously corresponds to the indicated strain ε̄SG and
can thus only be increased at the expense of reduced transducer stiffness kT :
1
βT ∝ ε̄SG ∝ (4.2)
kT
On the other hand, the strain is typically not distributed uniformly over the active
area of a strain gauge. As a result, the averaging over the strain-sensitive area A
causes a loss of sensitivity as indicated on the right of Figure 4.3:
Z
1
ε̄SG = ε dA . (4.3)
A B xx
All current Maltese cross sensors have shear beams with constant cross sections.
As shown by the following analysis, this design is not optimized for sensitivity
but mainly motivated by ease of design and manufacturing. The beam geometry
is therefore modified for high sensitivity by maximizing stress and achieving
constant strain in the active area of strain gauge installation. The “ideal” beam
geometry is examined using analytical beam theory and then verified numerically
by finite element analysis.
For a simplified analysis the Maltese cross-shaped transducer (cf. Figure 2.20)
is treated as a 2-D elastostatic mechanical system using shear beam theory. The
simplified model is shown on the left of Figure 4.4. Due to symmetry, the model
can be reduced to a single beam that extends along the x-axis. Without loss
of generality, we assume a square beam cross section of side bB . The beam is
supported by a prismatic bearing at its root (x = 0), allowing only rectilinear
translation along the z-axis: xu x=0 = 0. At x = l the beam is supported by a
membrane of thickness bM , height hM and length 2 · lM , which restrains the beam
end from rotation and transversal displacement. Provided that the membrane
is thin compared to the beam, blB bl M , the resisting moment caused by the
M
138 4 Sensor System
membrane is negligible. The membrane joint reduces to a pin connection with

2 u
deflection and moment fixed to zero: u|x=l = x2 = 0. The external force fz
x=l
is applied to the sensor hub in center of the transducer. Only the two beams
perpendicular to the load axis are deformed, hence the beams along the z-axis are
neglected in the reduced planar model. Then fz is equally spread on the x-axis
3
beams: Q = −EIyy xu3 = −fz /2. These assumptions finally lead to the simply
x=0
supported Euler-Bernoulli beam shown on the right of Figure 4.4. The surface
strain is given by
bB m
ε= . (4.4)
2EIyy
Here E is the Young’s modulus and m the moment

x

m = m0 1 − , with m0 = 21 fz l . (4.5)
l
Using this model, four different beam geometries with varying cross sections bB (x),
Figure 4.5, are compared in terms of their stiffness-sensitivity relationship:
1. Rectangular cross section

2. Cross section with constant surface strain
3. Trapezoidal cross section
4. Double-trapezoidal cross section
Rectangular Shaped Beam
The standard rectangular shaped beam (bB = const.) illustrated in Figure 4.5a is
considered for reference purposes. Assuming a square cross section of side bB , the
second moment of area is
bB4
Iyy = . (4.6)
12
Using Eqs. (4.5) and (4.6), the surface strain becomes
6m0 (1 − xl )
ε(a) = . (4.7)
EbB3
The elastic relations are normalized for better comparison of the different models.
For instance, Eq. (4.5) is being normalized to
m x
m0 = = 1 − x0 , with x0 = ∈ [0, 1] . (4.8)
m0 l
Here the prime (·)0 denotes unitless parameters. Normalizing Eq. (4.7) with
Eq. (4.8), E = 6 and bB = 1, we get the surface strain for the rectangular beam:
0
ε(a) = 1 − x0 . (4.9)
0
ε(a) 0
ε(b)
1.25 1.25
1.00 1.00
0.75 0.75
0.50 0.50
0.25 0.25
0 0
bB
b0
l l
(a) Rectangular cross section (b) Constant strain cross section
0
ε(c) 0
ε(d)
1.75 1.75
1.50 κ = 0.4 1.50 κ = 0.7
κ = 0.5 κ = 0.75
1.25 κ = 0.62 1.25 κ = 0.81
1.00 κ = 0.8 1.00 κ = 0.9
0.75 0.75
0.50 0.50
0.25 0.25
0 0
κ b0
κ b0
b0
b0
l xm
(c) Trapezoidal cross section (d) Double-trapezoidal cross section
Figure 4.5: Influence of shear beam geometry on surface strain and sensitivity
Cross Section with Constant Surface Strain

The averaging behavior of the strain gauges is eliminated if the beam cross section
is designed to satisfy the condition of constant surface strain along the beam
length as shown in Figure 4.5b. Solving Eq. (4.7) for bB results in the beam width
s s
x
6m 0 (1 − l ) 3 max(fz )l
bB = 3 , where b0 = bB |x=0 = 3 . (4.10)
Eε(b) Eε(b)
Here the indicated strain ε(b) is constant and chosen to match the properties of the
strain gauges. Furthermore, the initial beam width b0 is given by the maximum
load, max(fz ), and the beam length l and cannot be chosen arbitrarily. Normalizing
this equation using Eq. (4.8), E = 6 and ε(b) = 1 yields
√
3
bB0 = 1 − x0 . (4.11)
140 4 Sensor System
While this approach is advantageous to sensitivity, the curved surfaces resulting

from Eq. (4.11) are unfavorable for milling and strain gauge installation.
Trapezoidal Cross Section

The third approach is therefore based on a first-degree polynomial function. The
trapezoidal-shaped cross section shown in Figure 4.5c is a modification of the
rectangular beam. The beam width and height reduce arithmetically over the
beam length:
h i
bB = b0 · (κ − 1) xl + 1 . (4.12)
bB |x=l
Here b0 = bB |x=0 is the initial beam width and κ = b0 the slope ratio. Using this
equation in Eq. (4.4), we obtain
6m0 (1 − xl )
ε(c) = . (4.13)
Ebo3 [(κ − 1) xl + 1]3
Normalized with Eq. (4.8), b0 = 1 and E = 6, the above equation reduces to
0 1 − x0
ε(c) = . (4.14)
[(κ − 1) x0 + 1]3
0
The plot in Figure 4.5c shows the surface strain ε(c) for different slope ratios
κ. The sensitivity increases with decreasing κ but goes along with a significant
reduction in stiffness. A suitable trade-off between stiffness and sensitivity is
0
found for κ = 0.62: here ε(c) is flattened in the area of strain gauge installation
0
(0.1 ≤ x ≤ 0.3).
Double-Trapezoidal Cross Section

The main disadvantage of the previous design is the considerable loss of stiffness
over the beam length. The cross section is therefore modified in the fourth design
to achieve (almost) constant strain in the active area, while the loss of stiffness
is reduced. The idea is to introduce a weak spot in the area of strain gauge
installation. As in the previous design, the beam width bB reduces arithmetically
but increases immediately after the area of strain gauge installation. As a result,
the beam is composed of two pyramidal frustums with rectangular bases, adjoined
at their top bases. Figure 4.5d shows the double-trapezoidal cross section which is
defined by the piecewise function
κ−1 ·x+1 ,




 b 0 · xm 0 ≤ x ≤ xm ;
bB =  (4.15)


 b · 1 − κ · x + κl − xm , x < x ≤ l .
 0 l −x

l − xm m
m
Here bm = bB |x=xm = κ · b0 is the minimal beam width at x = xm and κ the slope

ratio. Assuming that bm is located in the middle of the beam (xm = 0.5), and using
Eq. (4.15) in Eq. (4.7), the surface strain becomes
6m0 (1 − xl )

3 , 0 ≤ x ≤ 2l ;



3 x

Eb 2(κ − 1) + 1


 0
 l
ε(d) =  x (4.16)
 6m (1
0 − l) l
, <x≤ .


 Eb3 2(1 − κ) x + 2κ − 1 3 2



0 l
Normalizing this equation with x0 = xl , E = 6 and b0 = M0 = 1 finally yields
1 − x0



 0 3 ; 0 ≤ x0 ≤ 0.5
 [2(κ − 1)x + 1]

0

ε(d) = (4.17)
 1 − x0 0
 [2(1 − κ)x0 + 2κ − 1]3 ; 0.5 < x ≤ 1



0
The plot in Figure 4.5d shows the surface strain ε(d) at different slope ratios κ.
As in the previous design, strain is increased by the diminished cross section for
x ≤ xm . Conversely, the extending cross section for x > xm limits stiffness losses.
A suitable trade-off between stiffness and sensitivity is found for κ = 0.81 where
0
ε(d) is flattened in the area of strain gauge installation (0.1 ≤ x0 ≤ 0.3). From the
manufacturing point of view, the double-trapezoidal cross section is noncritical.
In summary, the stiffness-sensitivity relationship is improved by using the
double-trapezoidal beam geometry. The tapered cross section acts as a notch that
causes a stress raise and consequently increases the strain level in the strain gauge
active area. Moreover, an almost uniform strain distribution in the area of strain
gauge installation is achieved by choosing an appropriate slope rate κ.
Verification by FEA
In the following the analytical results are verified numerically by using a de-
tailed finite element model in order to overcome idealized assumptions about
the analytical model and principal limitations of beam theory. The beam with
double-trapezoidal cross section is compared to the beam with rectangular shape,
Figures 4.5a and 4.5d, in terms of the stiffness-sensitivity relationship. Each
model comprises a single shear beam with its end supported by a membrane,
and parts of the inner sensor hub. Different load cases are applied to the sen-
sor hub using a compound of rigid elements. The model results are presented
in Figure 4.6. Each load case a–d corresponds to a single-axis force or moment
loading. The isosurfaces show the strain distribution for both models. In addition,
the strain |ε| along the normalized node path ξ 0 is plotted in the area of strain
gauge installation, and the integral strains |ε̄SG | indicated by the strain gauges
are labeled. Due to the qualitative nature of the results scaling is omitted for all
variables but all isosurfaces share the same color mapping. The results obtained
from FEA qualitatively correspond well with the analytical models. The strain
0 0
levels, ε(a) and ε(d) , Figure 4.5, are comparable to |ε| along node path ξ 0 . For the
double-trapezoidal beam a more “flattened” strain distribution in the strain gauge
active area is achieved by parameter modification. A good tradeoff is found for a
slope rate of κ = 0.85 (analytical result: κ = 0.81), as the notch stress concentration
142 4 Sensor System
Strain |ε(a) | [−]

Fy Fy |ε̄SG,a |
ξ ξ
0 0.1 0.2 0.3 0.4

Node path ξ 0 [−]
(a) Load case 1: F y
Strain |ε(b) | [−]

Fz Fz
ξ ξ
|ε̄SG,b |
0 0.1 0.2 0.3 0.4

(b) Load case 2: F z

Strain |ε(c) | [−]
|ε̄SG,c |
My ξ My ξ
0 0.1 0.2 0.3 0.4

(c) Load case 3: M y
|ε̄SG,d |
Strain |ε(c) | [−]
Mz Mz
ξ ξ
0 0.1 0.2 0.3 0.4

(d) Load case 4: M z
0 1 Qualitative strain |ε| Rectangular beam

Strain gauge active area Double-trapezoid beam
Figure 4.6: Numerical verification by FEA. The results for the rectangular and double-
trapezoid beams are shown qualitatively.
Table 4.1: Comparison of rectangular and double-trapezoidal beam geometry in terms

of stiffness and sensitivity
Load case
Criterion (a) fy (b) fz (c) my (d) mz
Stiffness K [–] 0.85 0.88 0.88 0.84

Sensitivity E [–] 1.38 1.44 1.44 1.40
Stiffness-sensitivity relationship S [–] 1.17 1.26 1.26 1.18
increases significantly for smaller κ.

Using the FEA results, Table 4.1 compares the stiffness of the rectangular and
double-trapezoid beam geometries by the ratio of axial displacement, u, and the
angle of twist, ϕ, for force and moment loading, respectively:
 u(a)
k(d) , force loading ;
 u(d)



KB = ϕ (4.18)
k(a)  ϕ (a) , moment loading .


(d)
Here, u and ϕ are calculated with respect to the point of force application. The
sensitivity of both designs is compared by the ratio of indicated strains:
|ε̄SG,d |
EB . (4.19)
|ε̄SG,a |
Finally, the stiffness-sensitivity relationship is introduced for better comparison

of the overall effect:
S B K·E . (4.20)
The indices given in Table 4.1 emphasize the superior performance of the novel
double-trapezoid beam geometry: although transducer stiffness is reduced by 12 to
16 %, sensitivity increases significantly by 38 to 44 %. Although the manufacturing
of the double-trapezoid shear beams is more involved, it is noncritical if a modern
CNC milling machine is used.
4.2.2 Mechanical Design

Based on the general requirements stated in Section 2.7.3 the mechanical system
of the sensor is developed. In this section, the overall design and assembly is
introduced, followed by a more detailed presentation of the transducer body. The
detailed FEA provides information on the strain level, stiffness and elastodynamic
properties.
Figure 4.7 shows photographs of the assembled F/T sensor. The top and bottom
covers are removed to reveal the strain gauges and sensor electronics. The sensor
has rather compact measurements (100 × 40 × 100 mm3 ). Weighing a total of 395 g,
the sensor is heavier than the one developed for the robot Johnnie (approximate
144 4 Sensor System
(a) Top view (b) Bottom view
Figure 4.7: Photographs of the six-axis force/torque sensor
weight 290 g). Still, the integration of the sensor as a load-bearing element within
the forefoot structure eliminates most of the surrounding foot structure and finally
reduces the overall weight of the foot segment.
The cross-sectional view in Figure 4.8a gives an overview of the sensor assembly.
The monolithic transducer body (1) is the key component of the sensor and is
described below in full detail. The bushes (2) are bonded to the inner flange of
the transducer and insulate the sensor from the rest of the robot. They are made
from phenolic cotton fabric laminate (HGW 2088) which shows high compressive
strength and superior insulation properties. The contact pressure under the bolt
heads (10) is spread over the lower insulation bush (2) by the load-spreading disk
(3). The screws (10) connect the sensor and ankle joint (not shown). Four overload
protection pins (4a) are mounted to the inner flange by press-fitting and secured
by set screws (4b). The functional principle of the overload protection mechanism
is described below. The transducer body provides seating surfaces for the sensor
electronics (5). The sensor cable (not shown) supplies electric energy and is used
for transmission of sensor data and configuration commands. It is secured by
a cable gland (not shown) to ensure reliable connections. Covers are integrated
on the top (7), bottom (8) and membranes (9) to protect the sensor against dam-
age and ingress of solid foreign objects and dust. The bottom cover (8) is made
from aluminum for higher flexural rigidity of the sensor assembly, whereas the
top and membrane covers (7,9) are made from plastic. The inside of the plastic
covers is coated with a copper-based, highly conductive coating to shield from
electromagnetic waves. Figure 4.8b shows the monolithic transducer body in a
standard Maltese-cross design with four equally spaced and crosswise-arranged
shear beams (3). The transducer is made from the high-strength aluminum alloy
EN AW-7075 T651 which offers high dimensional stability and typically very low
residual stress. The inner flange (1) connects to the ankle joint. Four bolted flange
connections (2) situated at the corners of the sensor frame interface to the forefoot.
The four shear beams (3) have the double-trapezoid shear beam cross section
developed in Section 4.2.1. The novel geometry of the sensing members signifi-
5
6 1 Transducer body
2 Insulation bush
3 Load-spreading disk
4a Overload protection pin
8 4b Set screw
1
5 Sensor electronics
6 Top cover
7 Bottom cover
8 Membrane cover
7 3 2 9 4b 4a 9 Ankle joint screw
(a) Cross-sectional view of sensor assembly
9
Detail X
X 10
3 4
3
Y
7
8
2
1
Detail Y
2
6
1 Ankle joint (inner) flange 4 Strain gauge 7, 8 Overload protection (bores)

2 Forefoot flange 5 Membrane 9 PCB seating surface
3 Double-trapezoid shear beam 6 Membrane slot 10 Cable bushing
(b) Design of transducer body and strain gauge placement
Figure 4.8: Mechanical design of the six-axis force/torque sensor
cantly improves the overall sensitivity, while the reduction of transducer stiffness
remains acceptable. Details X and Y in Figure 4.8b illustrate the orientation of the
strain gauges (4) on the shear beams (3). The meander-shaped measuring grid is
aligned with the beam longitudinal axis. The strain gauges (4) are bonded to the
center of the locating face using a hot-curing adhesive. Silicone rubber coating
provides mechanical protection of the gauges and soldering terminals. The mem-
branes (5) have a thickness of 1.3 mm and are cut from the transducer body by thin
slots (6) on the top and bottom side. The ends of the membrane slots are radiused
and the greater wall thickness reduces notch effects. The ankle joint connects
to the inner flange (1) at the center of the sensor. The bolted flange connection
is electrically isolated by bushes which are bonded to the transducer body and
isolate both flange and screws (see below). The outer sensor frame has four bolted
connections (2) with inclined flange faces to connect to the forefoot: two flanges
connect to the heel segment and the toe joint, respectively. Because of space and
weight restrictions the forefoot interface has a comparatively small flange area
and reduced screw count. To provide additional safety margin when transmitting
146 4 Sensor System
the high ground reaction forces EKagrip friction shims (cf. Section 3.2) are placed
between the connecting members to safely transmit the high loads over the com-
paratively small flanges. The transducer body comprises a mechanical overload
protection of the sensing members. Overload pins (not shown) are fixed to the
inner flange bore (7) and extend into a bore (8) of slightly larger diameter in the
outer sensor frame: the gap between the overload pin and the bore (8) determines
the permissible beam deflection. The gap size has been determined by FEA so that
the protection mechanism engages at a vertical load corresponding three times the
weight of the robot. If the predetermined load limit is exceeded, the overload pins
engage into the sensor load path and the additional “overload bypass” protects
strain gauges and sensor geometry. The complex geometry was machined from
the solid on a 5-axis CNC milling machine, removing more than 75 % of the total
volume of metal. Precision and tight tolerances were ensured by manufacturing
the beam and membrane geometry in only two settings from the top and bottom
side.
CATIA’s built-in FEA tool is used to iteratively refine the shear beam geometry
(dimensions, slope ratio, membrane shape, etc.). For a more thorough analysis
and fine tuning, a high-resolution finite element model is implemented in Altair
HyperWorks. The model setup shown in Figure 4.9 considers all components
within the sensor load path, including the transducer body, insulating bushes
and load-spreading disk. A hybrid mesh comprising tetrahedron and hexahedron
elements is created based on the CAD data. Good-quality hexahedral mesh is
chosen for the shear beams, while unstructured tetrahedral meshing is chosen for
other regions. This approach allows an accurate analysis of the sensor behavior
while the effort of mesh generation is kept manageable. Detail view X in Figure 4.9
illustrates the adaptive mesh refinement of the hexahedral mesh and the transition
to the unstructured mesh. The model is simply supported using appropriate
displacement boundary conditions for the forefoot flange surfaces. To ensure
homogenous and realistic load transmission to the sensor flange, the loads are
applied using a rigid body element and a reduced model of the ankle joint.
Figure 4.9 shows the results for different load cases that correspond to the maxi-
mum measurement range given in Eq. (4.1). Load case 3 (|my | = 120 Nm) yields
the highest stress: the maximum von Mises effective stress of approximately
140 N/mm2 occurs in the area of strain gauge installation and is below the threshold
of fatigue limit of the transducer material2 . Also, with |ε̄SG,my | = 1.424 · 10−3 the
indicated strain in this region is below the fatigue characteristics of the installed
strain gauges3 . From Figure 4.9 it can be seen that achieving high sensitivity both
for force and moment measurement are conflicting requirements: the sensor ge-
ometry is dimensioned for the lateral moments, mx and my , which finally reduces
the sensitivity of force measurement.
Table 4.2 specifies the translational and torsional stiffnesses of the transducer
body as calculated from the FEA results.
2 The dynamic fatigue strength of aluminum alloy EN AW-7075 T651 is rated at σFL =
150 N/mm2 [54].
3 The estimated fatigue life of the chosen strain gauges is rated at 107 cycles and a strain level
of ε̄SG = ±2.2 · 10−3 [194]
Detail X
X
(a) Model setup
(b) Load case 1: |fy | = 500 N (c) Load case 2: |fz | = 1200 N
(d) Load case 3: |my | = 120 Nm (e) Load case 4: |mz | = 50 Nm
Figure 4.9: FE analysis of the transducer body
Table 4.2: Calculated translational and torsional stiffnesses
Translational [N/m] Torsional [Nm/rad]
kx 3.675·107 kϕ,x 4.965·104

ky 3.650·107 kϕ,y 4.890·104
kz 5.089·107 kϕ,z 7.339·104
148 4 Sensor System
4.2.3 Sensor Electronics

The sensor employs metal foil strain gauges as sensing elements. They are chosen
to match the elastic properties of the transducer material to provide good self-
correction of temperature-dependent changes in the transducer sensitivity. Two
strain gauges are installed at opposing surfaces of each shear beam and connected
to half bridges for effective temperature compensation. The strain gauge place-
ment is shown in Figure 4.8b. An exact application was ensured by aligning the
strain gauges by a mechanical positioning jig. During setting of the adhesive, a
tailor-made clamping tool was used, providing constant pressure for constant glue
line thickness. The heat-curing adhesive required a two-stage curing process. At
any time the curing temperature was kept below the structural transformation
temperature of the precipitation hardening alloy EN AW-7075 T651. All these
measures finally lead to a high zero point stability of the sensor signals.
The half bridges are connected to the integrated sensor electronics. A simplified
circuit schematic is shown in Figure 4.10. The sensor electronics, Figure 4.10a,
comprise the signal conditioning and filtering circuits and A/D conversion for
eight strain gauge bridges. To ensure measurement accuracy and stability, a preci-
sion voltage source supplies the Wheatstone bridges illustrated in Figure 4.10b
and all other analog circuits. The bridge voltages are low-level and subject to noise
and offset errors. Before digitization, the signals are amplified and the offsets are
adjusted to match the available dynamic range of the A/D converter. Each channel
offset is adjusted individually by applying a programmable DC voltage from the
D/A converter to the reference pin of the instrumentation amplifier. The signals
are low-pass filtered to prevent aliasing effects in the digital signals and digitized
sequentially by the A/D converter. At a sampling rate of ≥8 kHz the A/D con-
verter has a resolution of ≥12 bit. The circuit contains a serial EEPROM for sensor
identification and storage of the calibration matrix to simplify sensor replacement
and make the signal processing software more portable. Signal processing runs on
an external DSP module, cf. Section 5.2. A serial interface (SPI) is used for data
exchange with the DSP module.
4.2.4 Force Sensor Calibration

The F/T sensor is designed for operation within the elastic range of the transducer
material. Hence there is a linear relationship between external forces and moments
and internal strains:
CS z = T lS . (4.21)
Here T l S = (T F T , T M T )T ∈ R6 is composed of the forces and moments acting on the

sensor with respect to sensor frame T , whose origin is in the center of the ankle
joint flange. z ∈ R8 is the vector of measured strains, and C S ∈ R6×8 the strain
compliance matrix to be identified. The design of the transducer body with four
equally-spaced, crosswise-arranged shear beams results in the selective sensitivity
of certain half bridges for certain forces or moments. Assuming the strain gauge
bridges are labeled from A to H, the theoretical configuration of the calibration
D
A
...
Uoff ,i
Wheatstone bridges (8×)
A
DSP module (external)

DAC
B
Multiplexer
A SPI
...
D
...
bG,i
U
Uref
ADC
H US Excitation
circuit EEPROM
(a) Schematic of sensor electronics
US
Uoff
UG bG
U
AGND
Amplifier Filter
(b) Single Wheatstone bridge
Figure 4.10: Simplified circuit schematic of the sensor electronics
matrix C S can be derived:
A B C D E F G H
 
 
 f 0 0 c1 0 0 0 −c1 0

 x 
 fy c2 0 0 0 −c2 0 0 0
 

 
C S =  fz  0 c3 0 c3 0 c3 0 c3 
 (4.22)
 mx 0 0 0 −c4 0 0 0 c4
 
 
 my 0 −c5 0 0 0 c5 0 0 


mz c6 0 c6 0 c6 0 c6 0

The coefficients ci (1 ≤ i ≤ 6) relate external loading with internal strains and

bridge voltages, respectively. Each line of the matrix shows the sensitivity of each
of the eight bridges to different basic loading conditions: for instance, moment
loading T M x causes a bending of the beams along the T y-axis, which is detected
by the bridges D and H. The quasi-symmetric design of the transducer yields
c1 ≈ c2 and c4 ≈ c5 . In practice, the configuration of the calibration matrix can
deviate for several reasons, including strain gauge misalignment errors, signal
noise and numerical errors during calibration.
Several methods have been proposed in the literature for the calibration of
multi-component force sensors. The most common ones are based on the least
squares method and perform with high accuracy, provided that the applied forces
150 4 Sensor System
are exactly known in magnitude and direction [42, 171, 198]. The set of forces
must be selected carefully, in particular,
– the number of applied forces k must at least equal the number of sensor
channels m; typically a redundant set is required (i. e., k m) to minimize
errors;
– the loads should be linearly independent;
– the loads should be of increasing magnitude, spanning the measurement
range in order to ensure accuracy over the entire working range.
Using Eq. (4.21), the series of measurement can be written in matrix notation as:
 T  T
z1  l 1 
 T   T 
z  l 2 
C S ·  .2  =  .  (4.23)
 ..   .. 
 T   T 
zk lk
Each line of the load matrix L ∈ R6×k and the measurement matrix Z ∈ R8×k holds
the applied loads l i ∈ R6 and the associated strain gauge values zi ∈ R8 of load
case i, respectively. The total number of load cases is k. Using Eq. (4.23), the
calibration matrix C S is computed using the Moore-Penrose generalized inverse
(pseudoinverse) of the measurement matrix Z :
CS = Z +L . (4.24)
The pseudoinverse Z + is then obtained accurately by singular value decomposi-

tion. Although mathematically simple and effective, least squares data fitting is
very sensitive to random noise such as errors in load magnitude and orientation,
measurement noise and numerical errors. On the other hand, it is well known that
calibration accuracy depends on the number of different loads that are applied.
Moreover, an accurate calibration requires the redundant measurement data to
achieve a better signal-to-noise ratio. The whole calibration procedure is thus very
time-consuming and costly, yet limited in accuracy.
The shape from motion approach proposed by Voyles et al. [195] originates from
machine vision. Calibration accuracy is comparable to least-squares methods, yet
this method reduces the effort of load application. As the above methods, it relies
on a high number of measurements but explicit knowledge of the applied loads
is not necessary. The raw data are continuously gathered while a proof-mass is
randomly moved by a manipulator. Only a small number of precisely known
loads are required to establish the sensor reference frame and to relate the raw
data. The calibration matrix is then recovered by singular value decomposition.
Load Application Procedure
The error of least squares calibration can only be minimized by exercising ex-
treme care during load application, thus, incorporating a redundant set of loads
is essential but very time-consuming. It is obvious to rationalize load applica-
tion by employing a systematic procedure and specialized calibration equipment.
Shimano and Roth [171] installed a force sensor on the wrist of an industrial
robot for in-situ calibration. The robot wrist was reoriented in various poses
5
5
B
1
4
1 Force/torque sensor
2 Swivel rotary table
3 Loading fixture
4 Adapter
5 Proof-mass
(a) Calibration setup
z
z
r y
x I
y
T
x
Configuration 1 Configuration 2 Configuration 3

(b) Proof-mass configurations
Figure 4.11: Setup for sensor calibration. Above, the loading fixture mounted on the
swivel rotary table of a 5-axis milling machine allows the precise application of a
highly redundant set of loads in a very efficient way; below, three different proof-
mass configurations are used to span the working range of the sensor.
with known proof-masses attached to the sensor, so that the loads could be easily
estimated. The accuracy of this procedure, however, was limited by the kinematic
and positioning accuracy of the robot arm. Uchiyama et al. [189] applied vari-
ous forces and moments by hanging known weights on a lever arm in different
configurations.
The load application procedure proposed here aims at applying and recording
a highly redundant set of load cases which should span the entire sensor working
range in a very accurate and time-effective way. Load application was conducted
in a semi-automated procedure on a 5-axis milling machine. High positioning
152 4 Sensor System
accuracy of the CNC machine enables precise and efficient application of a highly
redundant set of loads. Figure 4.11a shows the calibration setup. The F/T sensor
(1) is mounted on the swivel rotary table (2) of the milling machine by loading
fixture (3). Adapter (4) is bolted to the inner sensor flange. It provides five
mounting surfaces for the proof-masses (5) so that different load configurations
can be realized.
The procedure uses varying loads to determine the calibration matrix: three
different proof-mass configurations (Figure 4.11b) that differ in their CoM po-
sitions with respect to the sensor frame, span almost the entire sensor working
range. CoM positions are obtained from CAD data. For maximum accuracy all
components of the calibration setup were weighed and adjusted in the CAD model.
Depending on the proof-mass configuration, forces of approximately 500 N and
moments up to 80 Nm can be applied. The orientation of the force vector is mod-
ified incrementally by rotating the loading fixture about the T x and the T z axis
using the B-axis and C-axis of the swivel rotary table, respectively. Each sensor
was calibrated using 162 measured values per configuration, making a total of
486 load cases. The recording took approximately 3 hours per sensor.
Calibration and Verification

The high positioning accuracy of the CNC swivel rotary table and the knowledge
of all proof-mass CoMs allow the precise reconstruction of all load cases. Hence
the calibration matrix C S is obtained reliably from the least squares method. The
position vector T r = (rx , ry , rz )T from the sensor frame origin to the proof-mass
CoM is obtained from the validated CAD model (Figure 4.11 bottom). Using T r,
each line of the load matrix L in Eq. (4.23) can be written as

1 0 0 0 rz −ry 
!T !T 
F TFi

li = T i = = T F Ti · 0 1 0 −rz 0

rx  . (4.25)
 
T Mi Tr ×TFi
0 0 1 r −r 0 

y x
The force T F i acting on the sensor is obtained from the proof-mass weight ml :
TFi = ml · AT I · I g . (4.26)
Here I g = (0, 0, −g)T is the gravity vector with respect to the inertial frame I, AT I
the transformation matrix from the inertial into the sensor frame. AT I depends
on the spatial orientation of the CNC swivel rotary table which was recorded
automatically during the load application procedure. Using the so obtained
load matrix L and the measurement data matrix Z , the calibration matrix is
obtained from Eq. (4.24). Figure 4.12 shows the deviations of the measured values
T
T
l T = (F T , M T ) from the actual loads l̂ i = (b b Ti ). l i is calculated from Eq. (4.21),
Fi , M
i i i
while the actual loads l̂ i are obtained from Eq. (4.25). The errors of force and
moment measurement are defined as:
∆F i kF i − b
F ik ∆M i kM i − M
b ik
= ; = . (4.27)
Fi
b kbF ik M
bi kMb ik
4.3 Contact Sensors 153
7.0
6.5 k
+ Forces
6.0 Average forces (0.64 %)
Moments k
5.5
Measurement error [%] 5.0
Average moments (0.52 %)
4.5
4.0 + k
3.5
3.0 k
2.5
2.0 + + k
k
+ + k
1.5 ++ +++++++
k k k
+++ + ++ + +
k
++ + + ++ +++ ++++ k
++ k k + k
kk k k
k
1.0 + + k
+ ++ ++ kk k
+++ ++ ++++ + +++ +++ +++++ ++++++++ ++++++ +++
k
+ +++
k k k k k
++ + +++++ ++ ++ k k k k k
+ + + ++ + +++++++++++++++ ++++++++++++++++++++++ ++ +++ ++++++++ ++++ +++++++ ++++++++++++++++ ++ ++ ++ ++++ + +++ + + ++++++++ +++++++ +++
k k k k k k k
k k kkkkk kk k k
kk
k
k k k k k k
k
k k kk kk k
k
k k
kkk k k k
k k k k k k kkk k k k
k k k kkk kkk k k kkk kkkkkkkkk kk k
kk
0.5 ++++++++++++++++++++++++++++++ +++++++++ +++++++++++++++++ ++++++++++++++ +++++++++++ ++++++++++ + ++++++++ + kkk
k k k k k k k kk k k kkk k k k k kkk k k k k k kkk
k k k kkk k kk k kk kkkk k kk k k k k k k k k k k k k k
+++++++++ + ++++++++++ +++++++ +++++ +++++++++ kkk k kk k k k k kk kkk
k k
k k k k k
k kkk kk k k
k k k
k k kkkkk k k kkkkk kkkkk kk kk kkkk kk
k
k kkk kkkkk
k
kk k k k kkkk k k k kk k k kk k k k k kk
k kk kkkkkkkkkkkkkkkkkkk kk kkkkkkkkkkkkk kkkkkkkk kkk kkkkkkkkk k k k k kk kkk
kkk k k kk kk kkk k k k k k
kk k kk kk k kk
k k kk k kk
k kkkk k
k k kk kk k k kk kk k
k kk k k kk k k kk k kkk k k k k kkk k k k kkk k kkkk kk kkkkkk k
+ +++ ++ +++++++++ kkk k k+
kk kk kkk
++ + + + kk k
kk
+ + kk k
kk ++ ++ ++ ++ kk ++ k
k
k k k
k
k k kkk k k kkk
kk k k k
k kk k kk
0
k k kk
+
0 50 100 150 200 250 300 350 400 450 500
Load case
Figure 4.12: Verification of the least squares calibration
Deviations are calculated from the norm of vectorial difference between the calcu-
lated load and the theoretical wrench. In addition, the mean deviations provide
information on the calibration accuracy of the proposed method. The achieved
accuracy is 0.64 % and 0.52 % for force and moment measurement, respectively.
4.3 Contact Sensors

The segmentation of the foot into forefoot and hallux (cf. Section 3.9) yields an am-
biguous contact state of the foot segments. In order to detect and disambiguate the
contact state, four force-sensitive resistors are integrated into the contact elements
of each foot. The signals are digitized using Schmitt triggers and transmitted to a
local DSP module (see Section 5.2).
Knowledge of the contact state is necessary to calculate the ground reaction
forces: during terminal stance and the swing phase, only the toe segment has
ground contact but the force/torque sensor is located in the forefoot. Therefore, in
addition to the toe joint angle, measuring the contact state helps to reliably deter-
mine the ground reactions. Moreover, contact sensors enable the implementation
of event-triggered gait patterns.
4.4 Inertial Measurement System

The simulative study in Section 2.7.1 has shown that balance control is very
sensitive to measurement errors of the upper body orientation. The IMS is thus
one of the most critical components of the stabilizing system. It must show high
accuracy and signal quality and long-term stability. The overall bandwidth of the
sensor must be at least 40 Hz to 60 Hz.
Inertial measurement systems determine the orientation and angular velocity of
a rigid body from its rotational and translational movements. Today’s implemen-
tation of inertial measurement systems is in so-called strapdown systems where
three mutually orthogonal arranged gyroscopes and accelerometers, respectively,
154 4 Sensor System
Table 4.3: Technical data of high-accuracy inertial measurement systems
Criterion Unit iMAR iVRU-FC-C167 Crossbow VG700CB Crossbow IMU700CB
General
Update rate [Hz] 200 100 125
Weight [kg] 0.8 1.6 1.6
Angular rate
Sensor range [deg/s] ±200 ±200 ±200
Short-term bias [deg/s] ±0.003 ±0.03 ±0.03
Long-term bias [deg/h] ±36 ±20 ±20
Scale error [%] <0.2 <2 <2
Linearity [%] <0.2 <1 <1
Resolution [ /s]
deg <0.001 <0.0025 <0.025
Bandwidth [Hz] 200 >100 >100
√
Random walk [deg/ h] <0.02 <0.4 0.4
Linear acceleration
Sensor range [g] ±2 ±4 ±4
Short-term bias [mg] 0.5 n/a n/a
Long-term bias [mg] 2.5 <12 <12
Scale error [%] <0.3 <1 <1
Linearity [%] <0.3 <1 <1
Resolution [mg] <0.1 <0.5 <0.5
Bandwidth [Hz] >200 >10 >75
√
Random walk [m/s h] <0.02 <0.1 <0.1
are fixed to that rigid body. Due to accelerometer and gyro errors, the measured
angular rates and accelerations show an offset (bias) that results in integration
drift. Even small errors in measurement are integrated into progressively larger
errors in velocity and orientation. Integration drift can be compensated by mea-
suring other states of the system. Any sensor can be used that provides speed
or position information, such as a GPS receiver or magnetometer [56]. Similar to
human beings that utilize the eyes to supporting the sense of balance, the stereo
vision system of the robot can be utilized to provide the IMS with heading aiding
information [108, 152]. The inertial measurement system for the robot can be
realized by revising the customized IMS of the robot Johnnie [46] or by applying a
commercial system.
The revision of the Johnnie-IMS includes the redesign of all signal processing
electronics using simultaneous sampling of all sensors and oversampling. The
sensor fusion algorithms are extended by, e. g., error models, heading aiding from
optical flow. However, accelerometers and, particularly, gyroscopes with better
performance are barely available. Only a few systems for aerospace and military
applications were found. But these sensors are too expensive, compared to off-the-
shelf IMS. In summary, revising the Johnnie-IMS is a very elaborate task. The high
complexity and the tight project schedule implies that only existing technologies
could be applied without any new scientific results.
On the other hand, an extensive market survey has revealed that several com-
4.4 Inertial Measurement System 155
Figure 4.13: Open-frame inertial measurement system iVRU-FC-C167 from iMAR

Navigation
mercial IMS math the requirements of a humanoid robot. The technical data and
accuracy of inertial measurement systems of similar performance are listed in
Table 4.3. The most relevant criteria for comparison are:
Update rate: because the IMS is used within a closed control loop (sampling rate
approximately 1 kHz), the update rate should be as high as possible.
Accuracy (bias, linearity, random walk): the long-term stability (bias) of the
first system is lower than the other systems. However, comparing the other
accuracy data, the iVRU system is expected to be of better performance.
Weight: the weight of the sensor is another decisive criterion. The weight of the
first system is about 50 % of the other systems which is mainly due to the
customized open-frame design.
The high-precision inertial measurement system iVRU-FC-C167 from iMAR
Navigation [77] is built from fiber-optic gyroscopes and MEMS accelerometers.
Fiber-optic gyroscopes are distinguished from other gyros (like MEMS) in very low
noise, insensitivity against vibrations, high bandwidth, high zero-point stability,
and the absence of g-dependent drift. For example, the noise of a fiber-optic
gyro is one order of magnitude below the noise of a MEMS gyro. At an update
rate of 200 Hz the system comprises the sensor fusion algorithms and internal
error models. Also, it provides an open interface for heading aiding, e. g., by
a GPS receiver, magnetometer, or optic flow. The IMS comes in a customized
lightweight version weighing approximately 0.8 kg. A plastic cover protects the
system mechanically. Figure 4.13 shows a photograph of the open-frame inertial
measurement system iVRU-FC-C167 from iMAR Navigation.
Figure 4.14 shows the mounting arrangement for the inertial measurement
156 4 Sensor System
bb
1 Inertial measurement system

2 Protective cover
3 3 Connector
4 4 Status LEDs
7 5 Sensor carrier
5 6 IMS support structure
8 7 Trunk connection
8
8 Shock absorber
6 8 • CoM
Figure 4.14: Mounting arrangement for the inertial measurement system on the torso
system. The IMS (1) is mounted to the carrier (5) and protected by cover (2). The
connector (3) provides all the necessary interfaces for data output and configura-
tion. Two LEDs (4) indicate the status of the IMS. The support structure (6) which
holds the entire IMS assembly is mounted on the upper body using the clamping
(7). In the current setup, carrier (5) and support structure (6) are connected by
three pillars (not shown). The IMS is exposed to a rather high dynamic environ-
ment. In order to account for possible measurement errors, the support structure
is prepared for the installation of shock absorbers. A preliminary analysis of
walking on the spot showed peaks at 46 Hz, 55 Hz and above in the frequency
spectrum of the accelerometer and gyroscope signals. Since sensor fusion operates
at a sampling frequency of 100 Hz, a low-pass filtering of the raw sensor data with
a cutoff frequency of 50 Hz would be necessary to fulfill the sampling theorem.
According to signal processing theory, however, the accelerometer and gyroscope
signals cannot be filtered electronically. Therefore, it is be desirable to provide a
mechanism for vibration isolation that attenuates shock and vibrational energy
above 50 Hz. To this end, an isolation system can be interposed between the sensor
carrier (5) and the support structure (6). The system consists of three wire rope
isolators (8) that are capable of simultaneous shock and vibration attenuation in
all axes, have good damping properties and a fail-safe construction. The shock
absorbers (8) are arranged symmetrically with respect to the CoM of the sensor
assembly. The resultant force of each element passes through the sensor CoM
to avoid coupling modes. As a result of the external forces and torques, only
translational motion (x, y, z) of the IMS occurs, but rotation (from equilibrium)
about the coordinate axes (θx , θy , θz ) is omitted.
4.5 Chapter Summary

This chapter introduced the sensor system. It is able to acquire the complete
system state, except for the joint torques, and allows the implementation of model-
based control schemes:

Joint sensors: each joint contains an incremental rotary encoder, an absolute
angular sensor for link position sensing and a limit switch. The incremental
rotary encoder is mounted on the motor shaft and mainly used for motor
control. The absolute angular sensors on the output flanges allow direct
measurement of the joint angles, compensating elasticity and nonlinearity.
Furthermore, they enable the robot to start from arbitrary positions. For
operational security and to prevent the robot from self-destruction each joint
incorporates a limit switch.
Force/torque sensors: Each foot holds a custom-made six-axis force/torque sen-
sor that is tightly integrated to the foot structure. The sensor comprises a
monolithic transducer with four shear beams in a Maltese-cross arrangement.
Thin membranes mechanically decouple beam deflections and reduce cross
talk. The novel double-trapezoid beam geometry results in a uniform strain
distribution over the sensing area and increases overall sensitivity signifi-
cantly without reducing the transducer stiffness. An overload protection is
integrated to protect the sensor from damage: mechanical end-stops engage
into the force flux when exceeding a certain load and unload the sensitive
measurement beams.
Contact sensors: to detect and disambiguate the contact state of the feet, four
force-sensitive resistors are integrated to the contact elements of each foot.
Inertial measurement system: a high-precision inertial measurement system
that is built from fiber-optic gyroscopes and MEMS accelerometers is em-
ployed to estimate the upper body orientation and velocity with respect to
an inertial frame. The system is characterized by very low noise, insensi-
tivity against vibrations, high bandwidth, high zero-point stability, and the
absence of g-dependent drift.
5 Electronics Architecture
I’m completely operational and all my circuits are
functioning normally.
—HAL (2001, A Space Odyssey)
A robust and high-performance computer system is crucial to the performance

of the overall mechatronic system. This chapter briefly introduces the electron-
ics architecture and its essential components. Based on the considerations and
requirements developed in Section 2.8, the central control unit, local controllers
and the real-time communication system are presented. The chapter concludes
with an overview of the hardware security functions.
Electronics hardware development and programming are not part of this work:
the DSP board was developed by Mathias Bachmayer, all other electronics by
Georg Mayr. The hardware was tested and programmed by Valerio Favot.
As discussed in Section 2.8, the informational structure of the robot as an
“intelligent sensor-actuator network” accounts for reducing the overall complexity.
Figure 5.1 gives a schematic overview of the electronics architecture. Although
purely technologically motivated, similar hierarchical structures are found in bio-
logical systems, though of incomparably higher complexity. Lola is controlled by a
central control unit mounted on the upper body and nine local controllers that
are distributed over the robot. Fully decentralized control architectures are not
suitable because of the highly coupled kinematics and dynamics: stabilization of
global dynamics is therefore carried out by the central control unit. But low-level
control of the link positions and velocities is shifted into local controllers. Con-
cerning the electronics and control effort PMSM are considerably more elaborate
than BDCM. Thus, closing the low-level motor control loops co-located with the
actuators reduces the cabling effort and is expected to improve electromagnetic
compatibility. Each local controller can handle up to three servo systems. Actual
motor control is done on dedicated servo controller modules that are commanded
by the local controllers using the CANopen protocol. The local controllers also
account for sensor data acquisition and processing. In a sense of informational
pre-filtering only relevant data are transmitted to the higher-ranking control
layer. Gait generation and stabilization run fully autonomously on the on-board
computer system without any support from outside. An external PC is used only
for monitoring purposes and to give basic operating commands if the robot is
not connected to the vision system. Due to high computing power requirements,
vision processing is done on an external PC cluster.
Lola is mostly operated in the laboratory environment, so that energy is supplied
from an external source for convenience. There are separate power supplies for
the motors and electronics. The electronics power supply delivers a voltage of
48 V to the robot which is converted on-board into the voltages required for the
CCU and peripheral components. The motors are rated at 80 V, except for the
159
160 5 Electronics Architecture
Robot Central control unit Power supply

Electronic load
Local controller 1
80 V
Servo drive Servo controller a Motor power supply
LAN
CANopen
Servo drive Servo controller b 48 V Computer power
supply
Servo drive Servo controller c
SERCOS-III
Local controller i External processing
CANopen
Servo drive Servo controller a Monitoring PC
Servo drive Servo controller b
SPI
Ethernet switch
F/T sensor
Local controller 9 Vision processing

cluster
CANopen
Servo drive Servo controller a

CAN
Servo drive Servo controller b Camera 1

IMS Camera 2 GBit-LAN (2x)
Figure 5.1: Overview of the electronics architecture
miniature head actuators that are supplied with 24 V. Head motor supply voltage is
converted from the motor power supply. The motor controllers allow for operating
the motors in four quadrants. Thus, decelerating a motor can induce current into
the intermediate circuit, causing overvoltage. Therefore the motor power supply
consists of a power supply and an electronic load, acting as current source and
sink, respectively.
5.1 Central Control Unit

Trajectory generation and control of global system dynamics run on the CCU. A
PC-based platform is chosen because of the standardized hardware, convenient
programming and debugging, and high performance. It is based on an embedded
Mini-ITX mainboard with Intel Core 2 Duo Mobile processor (T7600, 2.33 GHz),
running the QNX real-time operating system. The concept of the “intelligent”
sensor-actuator network requires only an interface board to the real-time com-
munication system (Section 5.3) which is commercially available. An Ethernet
connection handles data exchange with the external monitoring PC.
5.2 Local Controllers

Nine local controllers carry out data acquisition, signal processing and low-level
motor control, unloading the CCU from these standard tasks. Off-the-shelf em-
5.2 Local Controllers 161
(a) Top view (b) Bottom view
Figure 5.2: Decentral control electronics composed of the DSP board, interface board
and two PCB-mounted Elmo servo controller modules
bedded systems in a small package that provide all necessary interfaces and
functionality are not available, thus, a proprietary development is necessary. The
small packaging goals and thus high integration density of the local controllers
requires to design multi-layer PCBs, using SMD-technologies with tiny chips like
ball grid arrays.
The design of the local controller modules follows a modular concept: the de-
central electronics are divided into a universal DSP board with an interface to the
real-time communication system. The DSP board is always used in combination
with an application-specific interface board that provides all the necessary inter-
faces and connectors to sensors, motor controllers, etc. Actual motor control is
done on dedicated servo controller modules by Elmo Motion Control [33] that are
commanded using the CANopen protocol. The Elmo modules are distinguished
by an extremely space-saving design and high power density due to the highly
efficient inverter. They comprise a fully digital motion controller that features cur-
rent, velocity and position loops, commutation alignment, gain scheduling modes
and built-in error protection. For most motors the PCB-mounted Whistle module
[33] is employed that weighs only 50 g and can support up to 20 A continuous and
40 A peak current. The maximum continuous power output is rated at 1.6 kW. The
hip joint flexion/extension and knee motors are driven by Guitar servo controllers
[33] that can deliver up to 25 A of continuous current and 50 A of peak current.
Weighing only 165 g, the maximum continuous power output is 2 kW.
Figure 5.2 shows a photograph of the local controller, the key features are listed
in Table 5.1. The DSP board is on top of the PCB stack and mounted piggyback to
the interface board. One local controller can connect up to three motor controllers:
two servo controllers are PCB-mounted to the bottom of the interface board and
fixed to the robot structure with their base plates for optimal heat transfer. The
third “satellite module” allows for tight integration of the controllers into the robot
structure and near the servo actuators. For example, the toe joint is controlled by
a satellite module housed on the shank, while the local controller is situated on
the thigh. Handling up to three motors by one controller is particularly useful
for multi-DoF joints because the inverse kinematics of the drive mechanism can
Table 5.1: Key features of the local controller hardware
DSP board
CPU DSP Freescale MC 56F8367
Communication interface SERCON 100S core on Xilinx FPGA XC3S400
Quadrature Decoder 2 channels
PWM 2×6 channels
Data bus 16 bit parallel bus (address range up to 16 MB)
CAN 2 ports
RS 232 2 ports
SPI 1 port for up to 5 slaves
Analog input 4 analog low pass filters up to fourth order
RAM Up to 8 MB SRAM mountable
Dimensions 80×55 mm2
Mounting Piggyback design
Interface board
Motor controllers 2 Elmo Whistle modules (PCB-mounted),
1 CAN port (satellite module)
RS 232 2 ports
Absolute angular sensors 3 EnDat IP cores on Xilinx FPGA XC3S400
F/T sensor 1 port (SPI)
Electronics power supply 12 V DC (plus on-board regulators)
Motor power supply 80 V DC
Dimensions 100×64.5 mm2
Mounting Piggyback design
be calculated locally, making the kinematic structure of the robot transparent to

higher control layers. For instance, the knee and ankle joints are handled by the
same local controller which allows to compensate knee motion in the ankle joint
mechanism (Section 2.6.3) without any cross-communication between the local
controllers.
5.3 Communication System

The real-time communication system handles the entire data exchange between
central controller running on the CCU and the local controllers. Being the only
data interface to the robot hardware, it must support the cyclic exchange of actual
and desired values, status and diagnostic messages, as well as the non-cyclic pa-
rameterization of sensors and drives. Besides the mandatory real-time capability,
the communication system must operate in a deterministic manner. The band-
width must be high enough to exchange all data throughout the system within one
communication cycle, corresponding to the sampling time of the CCU of approx-
imately 1 ms. The communication system should enable the testing of different
controller structures, ranging from more centralized to fully decentralized signal
processing concepts. In the first case only the current or velocity loop is closed in
the drive. All other loops are implemented in the CCU, thus the required commu-
5.3 Communication System 163
nication system bandwidth is considerably higher. The communication system

should use ring, line or bus topology to reduced the cabling effort; systems using
point-to-point cabling are not considered. For reasonable hardware development
effort a commercial interface board must be available for the CCU. The physical
interface of the local controllers should be available as ASICs or integrated into
the microprocessor.
A number of different real-time communication systems [12] were evaluated,
including CAN, FireWire, SERCOS and other proprietary implementations of
real-time Ethernet. As a result, a communication system based on SERCOS turned
out to be the most suitable solution.
Originating from the automotive sector, CAN-based communication systems
are widely used in industrial applications, including motion control. Advantages
of CAN are the easy setup and configuration of the bus structure and numerous
standard CAN-devices (sensors, controls and drives), PC interface cards and diag-
nostic tools. The purely event-triggered communication in one major drawback
of CAN, besides the rather low bandwidth of 1 Mbit/s. Hence peak loads can oc-
cur in cases of multiple simultaneous message transfer requests. However, such
requests occur frequently in multi-axis control applications. The mechanism of
bus arbitration ensures the priority-based transfer of all messages. But message
transfer is generally non-deterministic because it cannot be guaranteed within a
certain cycle. Using CAN in multi-axis control applications is therefore problem-
atic. Many controls overcome these shortcomings by reducing the communication
effort and longer cycle times. The trajectory generation is shared between the
CAN host and the servo nodes. However, such approaches are insufficient for the
impedance controlled ankle joints [18]. In this case long cycle times and trajectory
interpolation by the servo nodes is not applicable. Rather, updating the desired
position values within every cycle of the central control loop (approximately 1 ms)
is important to achieve stability.
Originally targeted at consumer-electronic applications, IEEE 1394 (FireWire) is
a serial bus with tree-like topology that connects nodes at data rates up to 800 Mbit/s.
It supports asynchronous and isochronous transactions. Isochronous data transfer
is used for data that require fixed, guaranteed bandwidth and deterministic
submission. IEEE 1394 can handle up to 63 isochronous channels. Isochronous
data are split into packets that are sent every 125 µs. Cycle Start Packets synchronize
the clock cycles of all nodes to this 8 kHz clock. Because isochronous transfer does
not retransmit lost or corrupted packets but delivers data at a constant rate, it is
well suited for time-triggered communication in distributed control systems [38].
Moreover, the entire communication system can be built from cheap standard
hardware and a variety of ASICs for interfacing the local controllers. Hardware
drivers, however, for connecting the local controllers are currently not available.
The in-house development is beyond the scope of this research.
The SERCOS interface [75] is an internationally standardized real-time commu-
nication system, targeting specifically at distributed motion control applications.
It supports cyclic exchange of actual and desired values with a high bandwidth.
Industrial applications include robotics, and tooling, paper and printing machines.
SERCOS uses ASICs as bus controllers and fiber optics as the transmission medium.
Synchronization between the central control unit and the drives is realized by
hardware with minimal jitter for accurate coordination of all drives. Likewise, all
actual values are acquired simultaneously and the desired values become valid
simultaneously. SERCOS communications uses Master Synchronization Telegrams
(MST), Master Data Telegrams (MDT), and Amplifier Telegrams (AT) to send
real-time information. For cyclic master-slave communication each drive (slave)
has its own time slot assigned to transfer data to the central controller (master).
SERCOS also supports configurations where multiple motors are assigned to a
single slave. Such setups are particularly useful for the parallel mechanisms of
the ankle joints. SERCOS includes a non-cyclic channel (service channel) to trans-
fer communication data, parameters and diagnostic data. The service channel
allows the parameterization of the drives, gain scheduling and switching between
different operating modes.
The current standard SERCOS-III [76] combines the communication and syn-
chronization mechanisms described before with Ethernet physics and twisted-pair
copper cable. In addition to the cyclic real-time and the non-cyclic service channel,
an IP channel can be added for transfer of non-real-time data without affecting
the real-time data exchange. Both real-time telegrams and IP channel telegrams
are embedded into standard Ethernet frames. SERCOS-III does not add a new
protocol to the standard Ethernet stack but puts Ethernet under control of the
motion bus in order to achieve a high level of determinism. Typical standard
Ethernet networks use the star topology with all devices connected to a central
switching device. As many other real-time Ethernet solutions SERCOS-III has line
or ring topology to reduce cabling. To permit these bus topologies with switched
Ethernet, a switch is needed in every connected node [36]. Due to the greater
bandwidth of Ethernet, SERCOS-III can connect a large number of slaves at cycle
times down to 31.25 µs. Thus, it is possible to implement both decentralized drive
concepts where all control loops are closed in the drive, and centralized signal
processing concepts where only the current loop is closed in the drive and all other
loops are implemented in the central controller. The open standard allows to add
data types for connecting the force/torque sensor and the inertial measurement
system to the communication system. To summarize, SERCOS-III provides the
following advantages:
– Time-slot mechanism for collision-free real-time communication
– High efficiency of the communication protocol
– Transfer of real-time and non-real-time data
– Identical topology, telegram structure and synchronization as former SER-
COS implementations
– Capable of cross-communication between the slaves
– Connects to proprietary devices by adding SERCOS data fields (force/torque
sensors, inertial measurement system)
At the time of writing (March 2010), the implementation of the SERCOS-III
communication system on the robot not finished yet. Instead, a CAN-based
system with eight parallel buses enabling point-to-point connections to the local
controllers, was used in order to speed up the first-time operation of Lola.
5.4 Security System 165
5.4 Security System

Regarding actuator performance and body size, the robot is a very dangerous
machine and can cause severe injury. In order to protect the operators, envi-
ronment and robot, several security functions are implemented. Effectiveness,
reliability and reaction time of software-based security functions is limited com-
pared to hardware-based systems. On the other hand, software-based solutions
can be triggered by certain states throughout the program flow. Moreover, cer-
tain functionality cannot be implemented in hardware: for example, the allowed
angular workspace of the two-DoF ankle joint mechanism (Section 2.6.3) is non-
rectangular. Thus, is cannot be limited by limit switches in the two rotational
axes but must rely on monitoring of the actual and desired values by software.
In conclusion, the combination of both software and hardware-based security
systems is the most reasonable approach. Possible hardware-triggered events
include:
– Motor overcurrent
– Over and undervoltage of the motor intermediate circuit voltage
– Motor and inverter over-temperature
– Angular range of the joints
In addition, there is an emergency stop circuit with several switches. In case
of failure, the operator can disable the motor power supply. CCU and local
controllers are supplied independently for simplified debugging.
6 Lessons Learned
To regret one’s own experiences is to arrest one’s own
development. To deny one’s own experiences is to put
a lie into the lips of one’s life. It is no less than a denial
of the soul.
—Oscar Wilde
In this chapter a series of recommendations for work related either specifically to

the humanoid robot Lola or other bipeds is given based on the author’s experiences
in the course of completing the research contained in this thesis.
6.1 Conceptual Issues
Body Size
Due to the tight schedule, it was not possible to investigate the influence of
body size on the walking performance. The proportions of Lola are thus equal to
the robot Johnnie. In order to make the robot more human-friendly, the author
proposes to consider a reduction in body size. Moreover, smaller robots are easier
to handle and less dangerous to the operators and experiments are simplified. On
the other hand, the stride length reduces with the leg length and leads to lower
walking speeds or higher joint velocities due to increased step frequencies.
Arm Movement
Leg advancement yields high moments about the contact normals which must be
compensated to avoid slippage. An effective counterforce is created by moving
other body parts opposed to leg advancement. Possible approaches are the upper
body rotation or arm swing. Arm movement is thus essential for fast locomotion.
The shoulder adduction/abduction joints and the newly introduced elbow joints
greatly improve the effectiveness of arm swing. In this configuration, however, the
movements of the arm endpoints are not fully decoupled from the upper body
orientation. The elbow joints permit translational movement of the arm endpoints
in the sagittal plane. In the transverse plane, however, mainly rotation about the
pelvis joint occurs, again yielding moments about the contact normals. Therefore,
an additional internal/external rotation joint in the upper arms is proposed to
enable translational arm movement in the transverse plane.
Knee Kinematics
Flexion/extension of the human knee is not a simple hinge movement, rather,
it performs a rotary-gliding motion: the instantaneous center of rotation moves
upward and backward during flexion. Prosthetic knee mechanisms try to resemble
human knee kinematics, for example by four-bar linkage mechanisms. Such
mechanisms are designed for minimal energy consumption during stance [148].
167
168 6 Lessons Learned
The angular range of the knee joint is sufficient for walking and running. For other
movements such as kneeling down, getting up or passing through narrow spaces
larger flexion angles would be useful. One approach to extend the motion range
would be to apply such a four-bar linkage mechanism to the knee joint. Moreover,
such a mechanism can further reduce actuator performance if designed for the
load profiles typical gait patterns. However, this mechanism cannot be combined
with the ankle joint drive mechanism developed in Section 3.8.3.
6.2 Mechanical System

The redundant kinematic structure with seven-DoF legs, the consistent lightweight
construction, a modular joint design with high power density and improved mass
distribution of the legs are the distinguishing characteristics of Lola’s mechanical
system.
Lightweight Design
Lightweight design and high effective stiffness of the mechanical structure obvi-
ously contribute to better robot performance. To this end, manufacturing major
structural components by investment casting has proven advantageous for achiev-
ing sophisticated designs with good dynamic characteristics and low weight. Due
to high manufacturing effort and cost, it is advisable to design these components
for high-cycle fatigue to avoid failure. Thus, the only way of reducing the weight
would be to reduce the loads that act on the robot, for instance by reducing the
body size or further reducing the moving masses. Alternatively, other materials
with higher specific stiffness could be used, e. g., fiber-reinforced polymers.
Actuator Performance
These considerations go along with lower actuator performance: lighter mechani-
cal structures yield smaller motors and transmissions and, thus, additional weight
savings. The servo actuators of Lola provide enough reserve capacity for imple-
menting various walking patterns. On the other hand, the performance of the
servo actuators is strictly constrained by the general weight restrictions. The
robot Lola provides a reliable platform for evaluating performance and overload-
ing capacity of the PMSM. The experimental results can help tailoring motor
performance to the actual needs of a next-generation robot.
Foot Dynamics
In the author’s opinion, refining on the mass distribution of the legs is by far more
significant. The evaluation of dynamic performance in Section 3.12.2 suggests that
actuator relocation is a suitable strategy to improve dynamic performance of the
system. The analysis also shows that the active toe joint contributes substantially to
resultant leg inertia. Despite the difficulties to finding suitable drive mechanisms,
actuator relocation of the active toe joint would considerably reduce leg inertia.
It is further proposed to compare active and passive toe joints regarding their
functionality in typical gait patterns. Passive toe joints should be chosen over
active ones if they are equal in effectiveness due to the obviously lower weight and
complexity.
6.3 Sensor System 169
Unfortunately, there is almost no relevant literature on the design of robotic feet.

Developing an ultra-lightweight foot concept that enables the robot to walk on
uneven terrain notwithstanding offers great potential for improving the walking
performance and further reduction of the resultant leg inertia.
6.3 Sensor System

With more than 50 sensors, the sensor layout of the robot Lola is rather complex.
For example, light barriers were applied to the robot Johnnie for angular referenc-
ing of the joints. For Lola, they are replaced by absolute angular sensors, enabling
direct measurement and control of the joint angles. Although the sensor count
remains constant, complexity of the electronics increases due to the proprietary
sensor interface. In summary, these sensors are nonetheless useful, especially with
regard to future applications in the human environment.
Joint Torque Measurement
The sensor layout of Lola is capable of acquiring the complete system state, except
for the joint torques. Integrating torque sensors to the leg joints—at least the
sagittal plane joints—would enable active compliant control and establish a basis
for evaluating more sophisticated, model-based control schemes. Thus, the robot
could be operated in position, velocity and torque mode. Active servo compliance
could help to attenuate the landing impacts more effectively. Moreover, active
control of the leg stiffness would allow the evaluation of biomechanical models
on a fully actuated humanoid robot. Another aspect of joint-level torque sensing
is related to safety. Lola is able to exert very high forces that can lead to severe
injuries. While the risk is controllable within the laboratory environment, addi-
tional measures are required when the robot is to interact with humans, sharing
the same workspace: for example, a collision detection based on joint-level torque
measurement would lead to much lower risk of injuries.
Various approaches to joint-level torque measurement for the different joint
types exist that differ in the modification effort. Rather simple solutions apply
to the knee and ankle joints: torque measurement in the knee joint is enabled
by applying strain gauges to the immersion tube (output link, see Figure 3.19)
to measure the tension and compression forces. Similarly, direct torque control
of the ankle joint is possible by integrating load cells into the coupler links (cf.
Figure 3.21). For example, calibrated off-the-shelf piezoelectric force links could
be applied. They are lightweight and compact, and very sensitive to tension
and compression forces. Applying torque sensors to the flexion/extension joint
of the hip and other Harmonic Drive-based rotary joints is more involved: it
can be accomplished by integrating force transducers into the output flange, or
measuring torsion of the gear housing.
Better Signal Quality and Accuracy
Local digitization of the force/torque sensor has reduced the measurement noise
significantly. To further improve signal quality and accuracy it is proposed to use
simultaneous sampling A/D converters for proper data synchronization. High
170 6 Lessons Learned
resolution sigma-delta (Σ-∆) converters should be used as they provide better

dynamic accuracy and eliminate phase errors. Σ-∆ converters allow the direct
digitization of the strain gauge bridges without the need for additional external
signal conditioning circuits.
6.4 Electronics Architecture

The main intention in the design of the electronics architecture was to unload the
central controller from hardware-related tasks such as data acquisition and motor
control. Moreover, less and smaller cable harnesses throughout the system were
expected by replacing the complex point-to-point cabling of sensors and actuators
by bus systems for communication and energy supply.
Cabling
All joint sensors are incorporated in the rotary servo actuators (Section 3.4.2).
Since the local controllers are co-located with the actuators a single cable harness
is needed to connect servo actuator and controller. For assembly convenience,
however, individual cables were used for the sensors and motor. In consequence,
the cabling of the robot more complex than expected due to the high number of
different sensors. It is therefore proposed to reduce the cable harnesses by using
one composite cable per joint. For instance, single strands could be combined
with a heat shrinkable tubing. This measure not only improves the appearance of
the robot, rather, it leads to weight savings and higher reliability. Furthermore, all
cables should be fitted with strain relief clamps as the miniature wire-to-board
connectors may easily loose contact due to vibrations and cable bending.
Local Controllers
Different interfaces for sensors and the real-time communication system made the
development of custom hardware necessary. The strict packaging goals involved
a substantial development effort to implement the decentral electronics. For
future developments, it is proposed to use off-the-shelf electronics hardware
whenever possible in order to achieve robust, EMI-resistant1 designs at reduced
development effort. In the meantime, several compact bus nodes for connecting
to real-time Ethernet became commercially available. Combined with a powerful
microprocessor module that incorporates all essential circuitry such as RAM,
Flash memory and standard communication interfaces, only an interface board
comprising application-specific peripherals must be developed.
1 Electromagnetic interference
7 Conclusions and Future Work
Wer am Ende ist, kann von vorn anfangen, denn das
Ende ist der Anfang von der anderen Seite.
—Karl Valentin
Humanoid robots are a new and promising application area for robotics. Bipedal
locomotion is considered one of the core technologies. All current full-size hu-
manoids are capable of stable dynamic walking but few are able to walk—or even
run—at speeds comparable to humans. Moreover, flexible motion generation in
realistic environments still remains challenging. Aiming at fast and autonomous
bipedal locomotion, the development of the humanoid walking robot Lola wishes
to make an important step towards this goal. Besides the challenging control
problems, fast bipedal locomotion puts high demands on the robot hardware—an
aspect which is often neglected or de-emphasized.
This thesis deals with the design and realization of the robot’s mechatronic
system. The issues covered include the mechatronic design concept, electrical
drive systems, design and analysis of the mechanical structure, sensor technology,
and contributions to the computer and electronics architecture.
The mechatronic system “humanoid walking robot” is characterized by a high
complexity. Here the systematic procedure adopted throughout the development
has proved to be successful. It is based on the virtual development of the complete
system prior to manufacturing actual hardware, using 3D-CAD and validation by
finite element analysis. The load profiles of the mechanical structure and actuators
are calculated with the comprehensive robot simulation model developed within
the sub-project “Simulation and Control.”
The robot has a kinematic structure with 25 actuated degrees of freedom in a
redundant configuration which allows more natural and flexible gait patterns. In
particular, actuated toe joints enable rolling the feet from heel to toe and reduce
the loads of other leg joints, compared to the standard leg configuration.
The mechanical structure is characterized by the consistent lightweight design
with high effective stiffness and natural frequencies. To this end, design proposals
for some components are created by topology optimization and the major links
are designed as aluminum investment castings in a Monocoque-like structure.
Numerical modal analysis is adopted to assess the dynamic properties of the
mechanical structure. The elastodynamic behavior of the drive mechanisms is
analyzed to avoid resonance oscillations and achieve high enough bandwidths and
fast response times of the speed control loops.
The joint drives form a basis for high dynamic performance and acceleration
behavior of the robot. The powerful yet lightweight servo actuators combine a
high-dynamic brushless servo motor with precision gearing and sensors into a
compact package. While most joints are driven through Harmonic Drive gears,
the knee and ankle joints employ linear actuators using planetary roller screws.
171
172 7 Conclusions and Future Work
Special attention is paid to the dynamic capabilities of the locomotor system.

The drive mechanisms of the knee and ankle joints place the actuators in the hip-
thigh area to minimize inertial contribution. Driving the ankle joint in a parallel
arrangement results in a more uniform load distribution among the actuators and
reduces the average motor loads significantly. As a result, superior acceleration
behavior and reduced effects of varying inertia are achieved, compared to other leg
architectures. In addition, the center of mass height is increased which reduces the
angular momentum about the fore-aft axis and, ultimately, enlarges the stability
margins available for balance control.
The dynamic performance is further affected by the kinematic constraints asso-
ciated with unilateral foot-ground contact. The foot dynamics concept therefore
ensures reliable contact on different surfaces. The feet further incorporate vis-
coelastic contact layers and passive heel segments with hydraulic shock absorbers
for effective impact attenuation at initial contact.
The sensor layout supports the implementation of model-based control algo-
rithms. Absolute angular sensors are employed for direct measurement of the
joint angles, compensating compliance and nonlinearity in the drive mechanisms.
The high-precision inertial measurement unit with fiber-optic gyroscopes and
MEMS accelerometers estimates the orientation and angular velocity of the upper
body. Furthermore, the ground reaction forces and moments are measured by
customized six-axis force/torque sensors.
The “intelligent” sensor-actuator network with central control unit improves
system performance from a technological point of view. The central control unit
accounts for trajectory planning and balance control. Low-level motor control
and sensor data acquisition and signal processing are done on nine decentral
controllers. The Ethernet-based real-time communication system SERCOS-III
handles the data exchange between all controllers with minimal latency.
Unfortunately, the early stage of experimental evaluation does not allow the final
assessment of the robot’s capabilities. First results nevertheless look very promis-
ing. At the time of writing, Lola is able to walk at a maximum of 3.34 km/h, which
makes it the fastest anthropomorphic humanoid walking robot. According to
publicly available data, Asimo achieves a walking (not running!) speed of 2.7 km/h,
all other bipeds are slower. Only PETMAN can walk as fast as 7.08 km/h, however,
this robot does not have anthropomorphic leg kinematics (cf. Section 1.1.1). Up to
now the locomotor system has proven to be extremely dynamic and seems to have
excellent acceleration behavior. These preliminary results support the hypothesis
that the proposed approach of low-inertia leg design provides significantly im-
proved agility, compared to traditional designs. Furthermore, the high quality
of the sensor signals is noteworthy. In particular, the balance controller benefits
from the very low noise and high accuracy of the inertial measurement system
and force/torque sensors. It is not easy to quantitatively assess the notion of
“human-like” or “natural” gait. In the author’s opinion, the redundant kinematic
structure with toe joints allows for a more human-like or natural looking gait. It
is left to the reader to view the videos of walking that are available on the project
website1 for an independent assessment.
1 http://www.amm.mw.tum.de/
173
The robot provides a reliable basis for the evaluation of different control ap-
proaches. In particular, the combination of vision system, mechatronics and
balance control give the robot a high degree of autonomy. At the HANNOVER
MESSE 2010, Lola was presented with a vision and path planning system de-
veloped by the Institute of Autonomous Systems Technology2 , University of the
Bundeswehr, München. The robot worked well and without considerable failure
during more than 25 demonstrations of vision-guided autonomous locomotion
and obstacle avoidance.
Future Work
Regarding the robot’s bipedal walking capabilities, the following recommenda-
tions for future work can be made.
The author believes that the further development of the foot dynamics concept,
Section 2.3, can make a significant contribution towards the goal of human-like
versatile locomotion.
Despite their importance for gait generation, the analysis in Section 3.12.2
has shown that the active toe joints counteract the low-inertia design of the legs.
Aiming at the development of ultra-lightweight feet, it is proposed to compare
active and passive toe joints with regard to their functionality in typical gait
patterns. Different concepts of passive toe joints should be evaluated, and passive
toe joints should be preferred if they are equal in effectivity to active toe joints.
The viscoelastic contact elements are crucial to the performance of ankle joint
torque control. Therefore, the systematic analysis and comparison of different
materials for shock attenuation and contact geometries is proposed. Finally, the
feet should be equipped with mechanisms for walking on uneven terrain.
Although reliable bipedal and autonomous locomotion are fundamental skills,
humanoid robots must have extended capabilities in order to be useful and com-
mercially successful. Possible extensions of the robot involve, for instance:
– The integration of manipulation capabilities (i. e., full arms and hands)
– Energy autarky
– Autonomous locomotion in more complex and realistic scenarios
– The implementation of safety features for human-robot interaction
2 http://www.unibw.de/lrt8/
Appendix A
Anatomic and Biomechanical
Fundamentals of Human Locomotion
Three basic reference planes are used in human anatomy that are shown on the
left of Figure A.1. Note that the same notation is used for planes that belong to the
total body or are fixed to a single link. The sagittal plane is the longitudinal plane
of symmetry (xz-plane) which extends from the front to the back of the body or
single link, dividing it into left and right portions. The sagittal plane runs through
the total or link CoM. The coronal plane is orthogonal to the sagittal plane and runs
through the CoM and divides the body or single link into anterior and posterior
sections (yz-plane). The transverse plane is orthogonal to the sagittal and coronal
plane and, running through the CoM, separates the body or a single link into
superior and inferior parts (xy-plane). The relative movement of body structures is
described with respect to these reference planes as shown on the right of Figure A.1.
The rotation of a link about an axis perpendicular to the coronal plane is called
flexion/extension. Flexion decreases the angle between two parts, and extension
increases it. For the ankle joint, these movements are called dorsiflexion and
plantarflexion; using the common notation for camera movement, these rotations
of the camera head are called tilt. Adduction and abduction describe the rotation of
a link about an axis that is orthogonal to the coronal plane. Adduction pulls a link
away from the midline, abduction pulls a structure or part towards the midline of
the body. Internal/external rotation describes the rotation of a link about an axis
perpendicular to the transverse plane. Internal and external rotation turn a link
inwards or outwards, respectively. For the camera head, this movement is called
pan.
175
176 Appendix A Anatomic and Biomechanical Fundamentals of Human Locomotion
Sagittal plane
Tilt
Coronal plane Pan
Transverse plane
Adduction/
abduction
Flexion/
extension
Internal/external
z rotation
Dorsiflexion/
plantarflexion
y
x
Adduction/
abduction
Figure A.1: Reference planes in human anatomy (left) and movement of body struc-
tures (right)
Appendix B
Main Dimensions
446
311
326
246
≈ 1727
1694
1467
440
430
60.5
114
130
≈ 446
177
Appendix C
Technical Data of the Joint Actuators
C.1 Mechanical
Load rating Gear
Drive Mass Motor inertia Top speed Continuous Peak Type Ratio
m [kg] Jm [kgm2 ] ϕ̇l,max [rad/s] τl,avg [Nm] τl,max [Nm] N [−]
Rotary actuators based on harmonic drive gears (cf. Section 3.4)

A 3.010 1.814·10−4 12.06 122∗ 243∗ CSG-32 50
−5
B 1.438 4.598·10 8.16 108 † 284† HFUC-25 100
−5
C 1.522 4.598·10 8.16 108 † 284† HFUC-25 100
−5
D 0.933 1.960·10 8.16 78 ∗ 147† HFUC-20 100
−5
E 1.082 2.375·10 8.16 78 ∗ 147† HFUC-20 100
−5
F 0.973 1.385·10 9.17 54 † 110† HFUC-17 100
−6
G 0.531 3.809·10 10.68 24 ∗ 48∗ CPL-14 100
Linear actuators based on planetary roller screws (cf. Sections 3.8.2 and 3.8.3)
Knee 1.931 5.544·10−5 10.68 ≤189∗ ≤378∗ RGTF-12×5 ≤72
−5‡
Ankle 0.516 1.544·10 10.68 ≤115 ∗ ≤230∗ RGTF-8×5 ≤70
Stereo camera head actuators (cf. Section 3.5)
Pan 0.237 5.00·10−7 8.90 0.90 7.50 CSF-11 100
Tilt 0.066 7.310·10−8 10.47 0.44 1.40† RSF-5A 100
Vergence 0.031 1.170·10−8 10.47 0.11 0.30† RSF-3B 100
∗ Limited by inverter current
† Limited by gear rating
‡ Only motor, motor-side transmission components are not considered
179
180 Appendix C Technical Data of the Joint Actuators
C.2 Electrical
Torque rating Constants
Drive Motor type∗ Continuous Peak Torque Motor Time

√
τm,c [Nm] τm,p [Nm] Kτ [Nm/A] Km [Nm/ W] Tel [ms]
Rotary actuators based on harmonic drive gears (cf. Section 3.4)

A PB K375100-4D 3.0 4.85† 0.097 0.267 1.53
B PB K064075-7Y 0.85 2.56 0.130 0.122 1.60
C PB K064075-7Y 0.85 2.56 0.130 0.122 1.60
D PB K064050-9D 0.62 1.56† 0.078 0.087 0.96
E PB K064025-GD 0.31 0.93 0.080 0.048 0.58
F PB K064025-GD 0.31 0.93 0.080 0.048 0.58
G PB K044025-JY 0.12 0.48 0.095 0.020 0.39
Linear actuators based on planetary roller screws (cf. Sections 3.8.2 and 3.8.3)
Knee PB K375050-5Y 1.72 5.14 0.105 0.153 1.45
Ankle PB K064050-8Y 0.62 1.87 0.109 0.087 1.15
Stereo camera head actuators (cf. Section 3.5)
Pan WI MSSI 022H-017C 0.025 0.09† 0.009 0.075 0.10
Tilt HD RSF-5A 0.004 0.014 0.011 0.008 0.32
Vergence HD RSF-3B 0.001 0.003 0.038 0.023 0.13
∗ PB = Parker Bayside [141]; HD = Harmonic Drive [60]; WI = Wittenstein [202].
† Limited by inverter current
Appendix D
Load Profiles
The reference gait pattern for the dimensioning of the mechanical system has been
determined by Buschmann using his multibody simulation model of the robot
(Section 3.3.1). The following gait parameters were used:
Walking speed 5.55 km/h

Step length 0.54 m
Step frequency 2.86 Hz
Double support phase 15.0 %
CoM (upper body) height 1.15 m
Shoulder flexion/extension (right arm)

6
0.9 5
4
Joint velocity ωl [rad/s]
0.8
Joint angle ϕl [rad]
3
0.7 2
0.6 1
0.5 0
0.4 -1
0.3 -2
-3
0.2 -4
0.1 -5
0 -6
11 12 13 14 15 16 17 11 12 13 14 15 16 17
40
30
20
Joint torque τl [Nm]
10
0
-10
-20
-30
-40
-50
-60
11 12 13 14 15 16 17
Time t [s]
181
182 Appendix D Load Profiles
Shoulder adduction/abduction (right arm)

-0.050 1.0
-0.075

0.5
-0.100
-0.125 0
-0.150
-0.5
-0.175
-0.200 -1.0
11 12 13 14 15 16 17 11 12 13 14 15 16 17
15
10
-5
-10
-15
11 12 13 14 15 16 17
Time t [s]
Elbow flexion/extension (right arm)

-0.9 5.0
2.5
-1.1
0
-1.3
-2.5
-1.5
-5.0
-1.7 -7.5
11 12 13 14 15 16 17 11 12 13 14 15 16 17
10
5
0
-5
-10
-15
-20
-25
-30
11 12 13 14 15 16 17
Time t [s]
183
Pelvis internal/external rotation

2.0
0.33 1.6

1.2
0.28
0.8
0.23 0.4
0.18 0
-0.4
0.13 -0.8
0.08 -1.2
-1.6
0.03
-2.0
-0.2 -2.4
11 12 13 14 15 16 17 11 12 13 14 15 16 17
75
50
25
-25
-50
-75
-100
11 12 13 14 15 16 17
Time t [s]
Pelvis adduction/abduction
0.12
1.5
0.08
1.0
0.04 0.5
0 0
-0.5
-0.04
-1.0
-0.08 -1.5
-0.12 -2
11 12 13 14 15 16 17 11 12 13 14 15 16 17
60
40
20
-20
-40
-60
11 12 13 14 15 16 17
Time t [s]
Hip internal/external rotation (right leg)

2.5
2.0

1.5
0.1
1.0
0.5
0
0
-0.5
-0.1 -1.0
-1.5
-0.2 -2
11 12 13 14 15 16 17 11 12 13 14 15 16 17
125
100
75
50
25
0
-25
-50
-75
-100
-125
11 12 13 14 15 16 17
Time t [s]
Hip adduction/abduction (right leg)

2.5
2.0
0.35 1.5
1.0
0.30
0.5
0
0.25
-0.5
0.20 -1.0
-1.5
0.15 -2
11 12 13 14 15 16 17 11 12 13 14 15 16 17
100
75
50
25
0
-25
-50
-75
-100
-125
-150
11 12 13 14 15 16 17
Time t [s]
185
Hip flexion/extension (right leg)

0.2 4

2
0
0
-0.2
-2
-0.4
-4
-0.6 -6
-0.8 -8
11 12 13 14 15 16 17 11 12 13 14 15 16 17
250
200
150
100
50
0
-50
-100
-150
-200
-250
-300
11 12 13 14 15 16 17
Time t [s]
Knee flexion/extension (right leg)

7.5
1.2
5.0
1.0 2.5
0
0.8
-2.5
0.6
-5.0
0.4 -7.5
11 12 13 14 15 16 17 11 12 13 14 15 16 17
150
100
50
0
-50
-100
-150
-200
-250
11 12 13 14 15 16 17
Time t [s]
Ankle adduction/abduction (right leg)

0.04 0.8
0.03 0.6

0.02
0.4
0.01
0.2
0
0
-0.01
-0.2
-0.02
-0.03 -0.4
-0.04 -0.6
-0.05 -0.8
11 12 13 14 15 16 17 11 12 13 14 15 16 17
60
40
20
-20
-40
-60
-80
11 12 13 14 15 16 17
Time t [s]
Ankle dorsiflexion/plantarflexion (right leg)

6
-0.1
4
-0.2
-0.3 2
-0.4 0
-0.5
-2
-0.6
-0.7 -4
11 12 13 14 15 16 17 11 12 13 14 15 16 17
175
150
125
100
75
50
25
0
-25
-50
-75
11 12 13 14 15 16 17
Time t [s]
187
Toe flexion/extension (right leg)

5
0
3

-0.1
1
-0.2
-1
-0.3 -3
-0.4 -5
-0.5 -7
-0.6 -9
11 12 13 14 15 16 17 11 12 13 14 15 16 17
30
25
20
15
10
5
0
-5
-10
-15
-20
-25
11 12 13 14 15 16 17
Time t [s]
Bibliography
[1] Aerotech, Inc. 1000 Series Brush, Rotary DC Servomotors. Product catalog. 2007. url:
http://www.aerotech.com/.
[2] Aerotech, Inc. S-Series Brushless, Frameless Torque Motors. Product catalog. 2007. url:
http://www.aerotech.com/.
[3] Akachi, K., Kaneko, K., Kanehira, N., Ota S. Miyamori, G., Hirata, M., Kajita, S.,
and Kanehiro, F. “Development of Humanoid Robot HRP-3P”. In: Proc. IEEE-RAS Int.
Conf. Humanoid Robot. (Humanoids). 2005, pp. 50–55. doi: 10.1109/ICHR.2005.1573544.
[4] Albers, A., Brudniok, S., Ottnad, J., Sauter, Ch., and Sedchaicharn, K. “ARMAR III
- Design of the upper body”. In: Proc. Workshop Human-Centered Robotic Systems (HCRS).
2006.
[5] Albert, A., Hofschulte, J., and Schermeier, O. “Entwicklung des zweibeinigen,
autonomen Laufroboters BARt-UH”. In: Proc. ROBOTIK. (in German). VDI-Verlag, 2000,
pp. 509–14.
[6] Angeles, J. and Park, F.C. “Performance Evaluation and Design Criteria”. In: Springer
Handbook of Robotics. Chap. 10, pp. 229–44.
[7] Asfour, T., Regenstein, K., Azad, P., Schröder, J., Bierbaum A. Vahrenkamp, N., and
Dillmann, R. “ARMAR-III: An Integrated Humanoid Platform for Sensory-Motor
Control”. In: Proc. IEEE-RAS Int. Conf. Humanoid Robot. (Humanoids). 2006, pp. 169–75.
doi: 10.1109/ICHR.2006.321380.
[8] Axsys Technologies, Inc. Brush Type DC Motors Handbook. San Diego CA, USA, 2005.
url: http://www.axsys.com/.
[9] Axsys Technologies, Inc. Brushless DC Motors Handbook. San Diego CA, USA, 2005.
url: http://www.axsys.com/.
[10] Baek, J.H., Kwak, Y.K., and Kim, S.H. “On the frequency bandwidth change of a servo
system with a gear reducer due to backlash and motor input voltage”. In: Arch. Appl. Mech.
73, 5-6 (2003), pp. 367–76. doi: 10.1007/s00419-003-0293-6.
[11] Beira, R., Lopes, M., Praca, M., Santos-Victor, J., Bernardino, A., Metta, G.,
Becchi, F., and Saltaren, R. “Design of the robot-cub (iCub) head”. In: Proc. IEEE Int.
Conf. Robot. Autom. (ICRA). 2006, pp. 94–100.
[12] Benzi, F., Buja, G.S., and Felser, M. “Communication architectures for electrical drives”.
In: IEEE Trans. Ind. Inf. 1, 1 (2005), pp. 47–53. doi: 10.1109/TII.2005.844428.
[13] Bevans, J.S. “Biomechanics: a review of foot function in gait”. In: The Foot 2, 2 (1992),
pp. 79–82. doi: 10.1016/0958-2592(92)90022-H.
[14] Bluethmann, W., Ambrose, R., Diftler, M., Askew, S., Huber, E., Goza, M.,
Rehnmark, F., C., Lovchik, and Magruder, D. “Robonaut: A Robot Designed to Work
with Humans in Space”. In: Auton. Robot. 14 (2003), pp. 179–97. doi:
10.1023/A:1022231703061.
[15] Bowling, A. “Mobility and Dynamic Performance of Legged Robots”. In: Proc. IEEE Int.
Conf. Robot. Autom. (ICRA). 2005, pp. 4100–7.
[16] Bruijn, S.M., Meijer O.G. van Dieën, J.H., Kingma, I., and Lamoth, C.J.C.
“Coordination of leg swing, thorax rotations, and pelvis rotations during gait: The
organisation of total body angular momentum”. In: Gait Posture 27, 3 (2008), pp. 455–62.
doi: 10.1016/j.gaitpost.2007.05.017.
189
190 Bibliography
[17] Buschmann, T., Lohmeier, S., Bachmayer, M., Ulbrich, H., and Pfeiffer, F. “A
Collocation Method for Real-Time Walking Pattern Generation”. In: Proc. IEEE-RAS Int.
Conf. Humanoid Robot. (Humanoids). 2007.
[18] Buschmann, T., Lohmeier, S., and Ulbrich, H. “Biped Walking Control Based on
Hybrid Position/Force Control”. In: Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS).
2009, pp. 3019–24. doi: 10.1109/IROS.2009.5354643.
[19] Buschmann, T., Lohmeier, S., Ulbrich, H., and Pfeiffer, F. “Dynamics Simulation for
a Biped Robot: Modeling and Experimental Verification”. In: Proc. IEEE Int. Conf. Robot.
Autom. (ICRA). 2006, pp. 2673–78.
[20] Buschmann, T., Lohmeier, S., Ulbrich, H., and Pfeiffer, F. “Optimization Based Gait
Pattern Generation for a Biped Robot”. In: Proc. IEEE-RAS Int. Conf. Humanoid Robot.
(Humanoids). 2005, pp. 98–103. doi: 10.1109/ICHR.2005.1573552.
[21] Carbone, G., Lim, H., Takanishi, A., and Ceccarelli, M. “Stiffness analysis of biped
humanoid robot WABIAN-RIV”. In: Mech. Mach. Theor. 41, 1 (2006), pp. 17–40. doi:
10.1016/j.mechmachtheory.2005.05.001.
[22] Carson, M.C., Harrington, M.E., Thompson, N., O’Connor, J.J., and Theologis, T.N.
“Kinematic analysis of a multi-segment foot model for research and clinical applications: a
repeatability analysis”. In: J. Biomech. 34, 10 (2001), pp. 1299–307. doi:
10.1016/S0021-9290(01)00101-4.
[23] Cheng, G., Hyon, Sang-Ho, Morimoto, J., Ude, A., Colvin, G., Scroggin, W., and
Jacobsen, S.C. “CB: A Humanoid Research Platform for Exploring NeuroScience”. In:
Proc. IEEE-RAS Int. Conf. Humanoid Robot. (Humanoids). 2006, pp. 182–87. doi:
10.1109/ICHR.2006.321382.
[24] ContiTech Antriebssysteme GmbH. CONTI SYNCHROBELT HTD Synchronous Drive
Belts. Product catalog. Hannover/Germany, 2002. url: http://www.contitech.de/.
[25] Croce, U.D., Riley, P.O., Lelas, J.L., and Kerrigan, D.C. “A refined view of the
determinants of gait”. In: Gait Posture 14, 2 (2001), pp. 79–84. doi:
10.1016/S0966-6362(01)00128-X.
[26] Cupec, R., Lorch, O., and Schmidt, G. “Vision-Guided Humanoid Walking – Concepts
and Experiments”. In: Proc. Int. Workshop Robotics in Alpe-Adria-Danube Region (RAAD).
2003.
[27] Dahmus, J.B., Gonzalez-Zugasti, J.P., and Otto, K.N. “Modular product architecture”.
In: Design Studies 22, 5 (2001), pp. 409–24. doi: 10.1016/S0142-694X(01)00004-7.
[28] Danaher Corp. Kollmorgen RBE(H) Series Motors. Wood Dale IL, USA, 2003. url:
http://www.danahermotion.com/.
[29] De Luca, A. and Book, W. “Robots with Flexible Elements”. In: Springer Handbook of
Robotics. Chap. 13, pp. 287–319.
[30] DIN 33402-1. Körpermaße des Menschen - Teil 1: Begriffe, Meßverfahren. (in German). Jan.
1978.
[31] DIN 33402-2. Ergonomie - Körpermaße des Menschen - Teil 2: Werte. (in German). Dec.
2005.
[32] Dresig, H. and Holzweißig, F. Maschinendynamik. 7th ed. (in German). Springer, 2006.
isbn: 3-540-22546-3.
[33] Elmo Motion Control, Ltd. SimplIQ Superior Servo Technology Products. Rev. 10.
Product catalog. Nov. 2009.
[34] ESK Ceramics GmbH & Co. KG. EKagrip Friction Shims. Technical data.
Kempten/Germany, 2008. url:
http://www.esk.com/fileadmin/esk/medien/pdf/TDS_ekagrip_e.pdf.
Bibliography 191
[35] Espiau, B. and Sardain, P. “The Anthropomorphic Biped Robot BIP2000”. In: Proc. IEEE
Int. Conf. Robot. Autom. (ICRA). 2000, pp. 3997–4002. doi: 10.1109/ROBOT.2000.845354.
[36] Felser, M. “Real-Time Ethernet – Industry Prospective”. In: Proc. IEEE 93, 6 (2005),
pp. 1118–29.
[37] Förg, D. and Ulbrich, H. “Simulation of a Biologically Inspired, Partly-Actuated Robot
Model with Passive Elements”. In: Proc. ECCOMAS Conf. Multibody Dynamics. 2009.
[38] Frommhagen, K., Nauber, P., Schelinski, U., and Scholles, M. “1394AP: a protocol
for deterministic industrial communication via IEEE 1394”. In: Proc. IEEE Conf. Emerging
Technologies and Factory Automation. 2005, pp. 933–40. doi:
10.1109/ETFA.2005.1612772.
[39] Fujiwara, K., Kanehiro, F., Kajita, S., Yokoi, K., Saito, H., Harada, K., Kaneko, K.,
and Hirukawa, H. “The first human-size humanoid that can fall over safely and stand-up
again”. In: Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS). 2003, pp. 1920–6.
[40] Fukaushima, T., Kuroki, Y., and Ishida, T. “Development of a new actuator for a small
biped walking entertainment robot”. In: Proc. Int. Conf. Power Electronics, Machines and
Drives (PEMD). 2004, pp. 126–31.
[41] Fukumoto, Y., Nishiwaki, K., Inaba, M., and Inoue, H.; “Hand-centered whole-body
motion control for a humanoid robot”. In: Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst.
(IROS). 2004, pp. 1186–91. doi: 10.1109/IROS.2004.1389557.
[42] Galway, R.D. A Comparison of Methods for Calibration and Use of Multi-component Strain
Gauge Wind Tunnel Balances. NRC/NAE Aeronautical Report LR-600. National Research
Council Of Canada Ottawa, Mar. 1980.
[43] Gard, S.A. and Childress, D.S. “The influence of stance-phase knee flexion on the
vertical displacement of the trunk during normal walking”. In: Arch. Phys. Med. Rehabil.
80, 1 (1999), pp. 26–32. doi: 10.1016/S0003-9993(99)90303-9.
[44] Gerecke, M., Hofschulte, J., and W., Gerth. “Realization of a Lightweight Sensory Foot
for a Bipedal Robot”. In: Proc. Int. Conf. Climbing and Walking Robots (CLAWAR). 2003,
pp. 895–902.
[45] Getzner Werkstoffe GmbH. Sylomer — cutting-edge material. 2009. url:
http://www.getzner.com/en/solutions/materials/sylomer/ (visited on
11/03/2009).
[46] Gienger, M. Entwurf und Realisierung einer zweibeinigen Laufmaschine. Fortschrittberichte
VDI, Reihe 1 378. (in German). Düsseldorf: VDI-Verlag, 2005. isbn: 3-18-337801-9.
[47] Gienger, M., Löffler, K., and Pfeiffer, F. “Towards the Design of a Biped Jogging
Robot”. In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 2001, pp. 4140–45. doi:
10.1109/ROBOT.2001.933265.
[48] Gienger, M., Löffler, K., and Pfeiffer, F. “Walking Control of a Biped Robot based on
Inertial Measurement”. In: Proc. Int. Workshop Humanoid Human Friendly Robot. (IARP).
2002, pp. 22–29.
[49] Gomi, H. and Hamaya, K. “Legged Walking Robot”. Pat. EP 1 097 860 (B1). Honda
Motor Co., Ltd. Oct. 12, 2003.
[50] Gomi, H., T., Kumagai, Hirose, M., and M., Nishikawa. “Articulated Structure for Legged
Walking Robot”. Pat. EP 0 433 096 (B1). Honda Motor Co., Ltd. Oct. 22, 1997.
[51] Gomi, H., Hamaya, K., Toyoda, H., and Takemura, Y. “Legged mobile robot”. Pat.
EP 1 535 705 (A1). Honda Motor Co., Ltd. Aug. 7, 2003.
[52] Gorinevski, D. M., Formalsky, A. M., and Schneider, A. Yu. Force Control of Robotics
Systems. CRC Press, 1997. isbn: 0-8493-2671-0.
192 Bibliography
[53] Gouaillier, D., Hugel, V., Blazevic, P., Kilner, C., Monceaux, J., Lafourcade, P.,
Marnier, B., Serre, J., and Maisonnier, B. “Mechatronic design of NAO humanoid”. In:
Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 2009, pp. 769–74.
[54] Haibach, E. Betriebsfestigkeit. 2. Aufl. (in German). Berlin ; Heidelberg [u.a.]: Springer,
2002. isbn: 3-540-43142-X.
[55] Hans Haugg Antriebstechnik GmbH & Co. KG. “Hermetic K-Series” bevel gear boxes.
Product catalog. Hiltenfingen/Germany, 2007.
[56] Harada, T., Sato, T., and Mori, T. “Development of Tiny Orientation Estimation Device
under Motion and Magnetic Disturbance”. In: Int. J. Robot. Res. 26, 5 (2007), pp. 547–59.
doi: 10.1177/0278364907079272.
[57] Harmonic Drive AG. CPL-2A Series Lightweight Component Sets. Product information.
Limburg/Lahn, 2007. url: http://www.harmonicdrive.de/.
[58] Harmonic Drive AG. CSF Mini Series Gearboxes. Product information. Limburg/Lahn,
2007. url: http://www.harmonicdrive.de/.
[59] Harmonic Drive AG. Precision in Motion. Full-line catalog. Limburg/Lahn, 2007/08.
url: http://www.harmonicdrive.de/.
[60] Harmonic Drive AG. RSF-Supermini AC Servo Actuator. Product information.
Limburg/Lahn, 2007. url: http://www.harmonicdrive.de/.
[61] Harzheim, L. Strukturoptimierung. 1. Aufl. (in German). Frankfurt am Main: Verlag Harri
Deutsch, 2008. isbn: 978-3-8171-1809-0.
[62] Hashimoto, K., Sugahara, Y., Hayashi, A., Kawase, M., Sawato, T., Endo, N.,
Ohta, A., Tanaka, C., and Takanishi, A. “New Foot System Adaptable to Convex and
Concave Surface”. In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 2007, pp. 1869–74. doi:
10.1109/ROBOT.2007.363594.
[63] Henze, A. “Dreidimensionale biomechanische Modellierung und die Entwicklung eines
Reglers zur Simulation zweibeinigen Gehens”. (in German). Ph.D. thesis. Tübingen:
Eberhard-Karls-Universität, 2002.
[64] Hirai, K., Hirose, M., and Takenaka, T. “The Development of Honda Humanoid Robot”.
In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 1998, pp. 1321–26. doi:
10.1109/ROBOT.1998.677288.
[65] Hirose, M. and Ogawa, K. “Honda humanoid robots development”. In: Phil. Trans. R. Soc.
A 365 (2007), pp. 11–19. doi: 10.1098/rsta.2006.1917.
[66] Hirzinger, G., Butterfass, J., Fischer, M., Grebenstein, M., Hahnle, M., Liu, H.,
Schaefer, I., and Sporer, N. “A mechatronics approach to the design of light-weight
arms and multifingered hands”. In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 2000,
pp. 46–54. doi: 10.1109/ROBOT.2000.844038.
[67] Hirzinger, G., Sporer, N., Albu-Schäffer, A., Hähnle, M., Krenn, R., Pascucci, A.,
and Schedl, M. “DLR’s torque-controlled light weight robot III – are we reaching the
technological limits now?” In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 2002,
pp. 1710–6. doi: 10.1109/ROBOT.2002.1014788.
[68] Hirzinger, G., Albu-Schäffer, A., Hähnle, M., Schaefer, I., and Sporer, N. “On a
New Generation of Torque Controlled Light-Weight Robots”. In: Proc. IEEE Int. Conf.
Robot. Autom. (ICRA). Vol. 4. Seoul, Korea, 2001, pp. 3356–3363. doi:
10.1109/ROBOT.2001.933136.
[69] Hobbelen, D., de Boer, T., and Wisse, M. “System overview of bipedal robots Flame and
TUlip: tailor-made for Limit Cycle Walking”. In: Proc. IEEE/RSJ Int. Conf. Intell. Robots
Syst. (IROS). 2008, pp. 2486–91. doi: 10.1109/IROS.2008.4650728.
Bibliography 193
[70] Hofschulte, J., Gerth, W., and Seebode, M. “Parallel Manipulator Hip Joint for a
Bipedal Robot”. In: Proc. Int. Conf. Climbing and Walking Robots (CLAWAR). 2004,
pp. 601–9. doi: 10.1007/3-540-29461-9_60.
[71] Hollerbach, J., Khalil, W., and Gautier, M. “Model Identification”. In: Springer
[72] Horakova, J. “The (Short) Robot Chronicle (On the 20th Century Cultural History of
Robots)”. In: Proc. Int. Workshop Robotics in the Alpe-Adria-Danube Region (RAAD). 2006.
[73] Hughes, J., Clark, P., and Klenerman, L. “The importance of the toes in walking”. In: J.
Bone Joint Surg. Br. 72, 2 (1990), pp. 245–51.
[74] Hurst, J.W., Chestnutt, J., and Rizzi, A. “Design and Philosophy of the BiMASC, a
Highly Dynamic Biped”. In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 2007, pp. 1863–8.
doi: 10.1109/ROBOT.2007.363593.
[75] IEC/EN 61491. Electrical equipment of industrial machines – Serial data link for real-time
communication between controls and drives. Oct. 2002.
[76] IEC/PAS 62410. Real-time Ethernet SERCOS III. Aug. 2005.
[77] iMAR GmbH. iVRU-FC: Vertical Reference Unit with Fiber Optic Gyros, MEMS
Accelerometers and integrated Strapdown Processor. Technical Datasheet. St.
Ingbert/Germany, 2007.
[78] INA Schaeffler KG. Planetary roller screws and bearing components (RGT). Product
catalog. Homburg/Germany, 2000. url: http://www.ina.com/.
[79] INA Schaeffler KG. Wälzlager. Product catalog (in German). Herzogenaurach/Germany,
2006. url: http://www.ina.com.
[80] International Federation of Robotics (IFR), ed. World Robotics 2008. VDMA Verlag,
2008.
[81] Isermann, R. Mechatronische Systeme – Grundlagen. (in German). Springer, 1999. isbn:
3-540-64725-2.
[82] Ishida, T. and Takanishi, A. “A Robot Actuator Development With High
Backdrivability”. In: Proc. IEEE Conf. Robotics, Automation and Mechatronics. 2006, pp. 1–6.
doi: 10.1109/RAMECH.2006.252631.
[83] Ishida, T., Kurokoi, Y., Nagasaka, K., and Yamaguchi, J. “A Small Biped
Entertainment Robot and its Attractive Applications”. In: Proc. Int. Workshop Humanoid
Human Friendly Robot. (IARP). 2002, pp. 116–9.
[84] Ishida, T., Kuroki, Y., and Yamaguchi, J. “Mechanical system of a small biped
entertainment robot”. In: Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS). 2003,
pp. 1129–34. doi: 10.1109/IROS.2003.1248796.
[85] Kajita, S. and Espiau, B. “Legged Robots”. In: Springer Handbook of Robotics. Chap. 16,
pp. 361–89.
[86] Kajita, S. and Tani, K. “Adaptive Gait Control of a Biped Robot Based on Realtime
Sensing of the Ground Profile”. In: Auton. Robot. 4, 3 (1997), pp. 297–305. doi:
10.1023/A:1008848227206.
[87] Kajita, S., Nagasaki, T., Kaneko, K., Yokoi, K., and Tanie, K. “A hop towards running
humanoid biped”. In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 2004, pp. 629–35. doi:
10.1109/ROBOT.2004.1307219.
[88] Kajita, S., Kanehiro, F., Kaneko, K., Fujiwara, K., Yokoi, K., and Hirukawa, H. “A
Realtime Pattern Generator for Biped Walking”. In: Proc. IEEE Int. Conf. Robot. Autom.
(ICRA). 2002, pp. 31–7. doi: 10.1109/ROBOT.2002.1013335.
[89] Kajita, S., Nagasaki, T., Kaneko, K., Yokoi, K., and Tanie, K. “A Running Controller of
Humanoid Biped HRP-2LR”. In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 2005,
pp. 616–22.
194 Bibliography
[90] Kajita, S., Kaneko, K., Morisawa, M., Nakaoka, S., and Hirukawa, H. “ZMP-based
Biped Running Enhanced by Toe Springs”. In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA).
2007, pp. 3963–9. doi: 10.1109/ROBOT.2007.364087.
[91] Kaminaga, H., Ono, J., Nakashima, A., and Nakamura, Y. “Development of
Backdrivable Hydraulic Joint Mechanism for Knee Joint of Humanoid Robots”. In: Proc.
IEEE Int. Conf. Robot. Autom. (ICRA). 2009, pp. 1577–82.
[92] Kanehiro, F., Ishiwata, Y., Saito, H., Akachi, K., Miyamori, G., Isozumi, T.,
Kaneko, K., and Hirukawa, H. “Distributed Control System of Humanoid Robots based
on Real-time Ethernet”. In: Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS). 2006,
pp. 2471–77. doi: 10.1109/IROS.2006.281691.
[93] Kaneko, K., Kanehiro, F., Morisawa, M., Miura, K., Nakaoka, S., and Kajita, S.
“Cybernetic Human HRP-4C”. In: Proc. IEEE-RAS Int. Conf. Humanoid Robot. (Humanoids).
2009, pp. 7–14. doi: 10.1109/ICHR.2009.5379537.
[94] Kaneko, K., Kajita, S., Kanehiro, F., Yokoi, K., Fujiwara, K., Hirukawa, H.,
Kawasaki, T., Hirata, M., and Isozumi, T. “Design of Advanced Leg Module for
Humanoid Robotics Project of METI”. In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 2002,
pp. 38–45. doi: 10.1109/ROBOT.2002.1013336.
[95] Kaneko, K., Kanehiro, F., Kajita, S., Hirukawa, H., Kawasaki, T., Hirata, M.,
Akachi, K., and Isozumi, T. “Humanoid Robot HRP-2”. In: Proc. IEEE Int. Conf. Robot.
Autom. (ICRA). 2004, pp. 1083–90. doi: 10.1109/ROBOT.2004.1307969.
[96] Kaneko, K., Harada, K., Kanehiro, F., Miyamori, G., and Akachi, K. “Humanoid
robot HRP-3”. In: Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS). 2008, pp. 2471–78.
doi: 10.1109/IROS.2008.4650604.
[97] Kemp, C.C., Fitzpatrick, P., Hirukawa, H., Yokoi, K., Harada, K., and Matsumoto, Y.
“Humanoids”. In: Springer Handbook of Robotics. Chap. 56, pp. 1307–33.
[98] Kerrigan, D.C., Croce, U.D., Marciello, M., and Riley, P.O. “A refined view of the
determinants of gait: significance of heel rise”. In: Arch. Phys. Med. Rehabil. 81 (2000),
pp. 1077–80. doi: 10.1053/apmr.2000.6306.
[99] Kim, G., Shin, H., and Yoon, J. “Development of 6-axis force/moment sensor for a
humanoid robot’s intelligent foot”. In: Sensor Actuator Phys. 141, 2 (2008), pp. 46–53. doi:
j.sna.2007.08.011.
[100] Kim, J., Park, I., Lee, J., Kim, M., Cho, B., and Oh, J. “System Design And Dynamic
Walking Of Humanoid Robot KHR-2”. In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 2005,
pp. 1443–48.
[101] Klein, B. Leichtbau-Konstruktion. 4th ed. (in German). Vieweg, 2000. isbn: 3-528-34115-7.
[102] Konno, A., Sellaouti, R., Amar, F.B., and Ouezdou, F.B. “Design and development of
the biped prototype ROBIAN”. In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 2002,
pp. 1384–9. doi: 10.1109/ROBOT.2002.1014736.
[103] Krishnan, R. “Selection Criteria for Servo Motor Drives”. In: IEEE Trans. Ind. Applicat.
23, 2 (1987), pp. 270–5. doi: 10.1109/TIA.1987.4504902.
[104] Kwon, W., Kim, H.K., Park, J.K., Roh, C.H., Lee, J., Park, J., Kim, W., and Roh, K.
“Biped Humanoid Robot Mahru III”. In: Proc. IEEE-RAS Int. Conf. Humanoid Robot.
(Humanoids). 2007.
[105] Li, J., Huang, Q., Zhang, W., Yu, Z., and Li, K. “Flexible foot design for a humanoid
robot”. In: Proc. IEEE Int. Conf. Automation and Logistics (ICAL). 2008, pp. 1414–9. doi:
10.1109/ICAL.2008.4636375.
[106] Lim, H. and Takanishi, A. “Biped walking robots created at Waseda University: WL and
WABIAN family”. In: Phil. Trans. R. Soc. A 365, 1850 (2007), pp. 49–64. doi:
10.1098/rsta.2006.1920.
Bibliography 195
[107] Lindemann, U. Methodische Entwicklung technischer Produkte. 2nd ed. (in German).
Springer, 2007. isbn: 978-3-540-37435-0.
[108] Lobo, J. and Dias, J. “Vision and inertial sensor cooperation using gravity as a vertical
reference”. In: IEEE Trans. Pattern Anal. Machine Intell. 25, 12 (2003), pp. 1597–1608. doi:
10.1109/TPAMI.2003.1251152.
[109] Löffler, K. Dynamik und Regelung einer zweibeinigen Laufmaschine. Fortschrittberichte
VDI, Reihe 8 1094. (in German). Düsseldorf: VDI-Verlag, 2006. isbn: 3-18-509408-5.
[110] Löffler, K., Gienger, M., Pfeiffer, F., and Ulbrich, H. “Sensors and control concept of
a biped robot”. In: IEEE Trans. Ind. Electron. 51, 5 (2004), pp. 972–80. doi:
10.1109/TIE.2004.834948.
[111] Lohmeier, S., Buschmann, T., Ulbrich, H., and Pfeiffer, F. “Design Concept of
Humanoid Robot LOLA”. In: Proc. CISM-IFToMM Symposium on Robot Design, Dynamics,
and Control (RoManSy). 2008.
[112] Lohmeier, S., Buschmann, T., Ulbrich, H., and Pfeiffer, F. “Humanoid Robot LOLA —
Research Platform for High-speed Walking”. In: Motion and Vibration Control — Selected
Papers from MOVIC 2008. Ed. by Ulbrich, H. and Ginzinger, L. Springer, 2008,
pp. 221–30. isbn: 978-1-4020-9437-8.
[113] Lohmeier, S., Buschmann, T., Schwienbacher, M., Ulbrich, H., and Pfeiffer, F. “Leg
Design for a Humanoid Walking Robot”. In: Proc. IEEE-RAS Int. Conf. Humanoid Robot.
[114] Lohmeier, S., Buschmann, T., and Ulbrich, H. “System Design and Control of
Anthropomorphic Walking Robot LOLA”. In: IEEE/ASME Trans. Mechatron. 14, 6 (2009).
Focused Section on Anthropomorphism in Mechatronic Systems, pp. 658–66. doi:
10.1109/TMECH.2009.2032079.
[115] Maxon Motor AG. Full-line Catalog. Sachseln (Switzerland), 2008. url:
http://www.maxonmotor.com/.
[116] McGeer, T. “Passive Dynamic Walking”. In: Int. J. Robot. Res. 9, 2 (1990), pp. 62–82. doi:
10.1177/027836499000900206.
[117] Miyachi Systems Corp. Miniature Humanoid Robots. (in Japanese). 2010. url:
http://www.miyachi-sys.com/services/robot/index.html (visited on 02/15/2010).
[118] Miyazaki, S. and Otani, T. “Leg Type Movable Robot”. Pat. EP 1 609 567 (A1). Honda
Motor Co., Ltd. Dec. 28, 2005.
[119] Miyazaki, S. and Takahashi, H. “Leg Structure of Legged Robot”. Pat. EP 1 344 614 (A1).
Honda Motor Co., Ltd. Sept. 17, 2003.
[120] Murata, S., Yoshida, E., Tomita, K., Kurokawa, H., Kamimura, A., and Kokaji, S.
“Hardware design of modular robotic system”. In: Proc. IEEE/RSJ Int. Conf. Intell. Robots
[121] Nagakubo, A., Kuniyoshi, Y., and Cheng, G. “Development of a high-performance
upper-body humanoid system”. In: Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS).
2000, pp. 1577–1583. doi: 10.1109/IROS.2000.895198.
[122] Nagakubo, A., Kuniyoshi, Y., and Cheng, G. “ETL-Humanoid – A high-performance
full body humanoid system for versatile actions”. In: Proc. IEEE/RSJ Int. Conf. Intell. Robots
[123] Nakamura, Y., Okada, M., Shinohara, T., and Goto, T. “Mechanical Challenges for
Further Humanoid Robot Evolution”. In: Proc. Third IARP International Workshop on
Humanoid and Human Friendly Robotics. 2002.
[124] Nakanishi, Y., Namiki, Y., Hongo, K., Urata, J., Mizuuchi, I., and Inaba, M. “Design
of the musculoskeletal trunk and realization of powerful motions using spines”. In: Proc.
IEEE-RAS Int. Conf. Humanoid Robot. (Humanoids). 2007, pp. 96–101. doi:
10.1109/ICHR.2007.4813854.
196 Bibliography
[125] Nishiwaki, K., Murakami, Y., Kagami, S., Kuniyoshi, Y., Inaba, M., and Inoue, H. “A
Six-axis Force Sensor with Parallel Support Mechanism to Measure the Ground Reaction
Force of Humanoid Robot”. In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 2002,
pp. 2277–82. doi: 10.1109/ROBOT.2002.1013571.
[126] Nishiwaki, K., Kagami, S., Kuffner, J., Inaba, M., and Inoue, H. “Humanoid ‘JSK-H7’:
Research Platform for Autonomous Behavior and Whole Body Motion”. In: Proc. Int.
Workshop Humanoid Human Friendly Robot. (IARP). 2002, pp. 2–9.
[127] Nishiwaki, K., Kuffner, J., Kagami, S., Inaba, M., and Inoue, H. “The experimental
humanoid robot H7: a research platform for autonomous behaviour”. In: Phil. Trans. R.
Soc. A 365, 1850 (2007), pp. 79–107. doi: 10.1098/rsta.2006.1921.
[128] Nishiwaki, K., Kagami, S., Kuniyoshi, Y., Inaba, M., and Inoue, H. “Toe Joints that
Enhance Bipedal and Fullbody Motion of Humanoid Robots”. In: Proc. IEEE Int. Conf.
Robot. Autom. (ICRA). 2002, pp. 3105–10. doi: 10.1109/ROBOT.2002.1013704.
[129] Novacheck, T.F. “The biomechanics of running”. In: Gait Posture 7 (1998), pp. 77–95. doi:
10.1016/S0966-6362(97)00038-6.
[130] Ogura, Y., Aikawa, H., Lim, H.-O., and Takanishi, A. “Development of a human-like
walking robot having two 7-DOF legs and a 2-DOF waist”. In: Proc. IEEE Int. Conf. Robot.
[131] Ogura, Y., Aikawa, H., Shimomura, K., Kondo, H., Morishima, A., Lim, H., and
Takanishi, A. “Development of A Humanoid Robot WABIAN-2”. In: Proc. IEEE Int. Conf.
Robot. Autom. (ICRA). 2006, pp. 76–81.
[132] Ogura, Y., Shimomura, K., Kondo, H., Morishima, A., Okubo, T., Momoki, S., Lim, H.,
and Takanishi, A. “Human-like Walking with Knee Stretched, Heel-contact and Toe-off
Motion by a Humanoid Robot”. In: Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS).
2006, pp. 3976–81. doi: 10.1109/IROS.2006.281834.
[133] Okada, M. and Nakamura, J. “Development of a Cybernetic Shoulder – A Three-DOF
Mechanism That Imitates Biological Shoulder Motion”. In: IEEE Trans. Robot. 21, 3 (2005),
pp. 438–43. doi: 10.1109/TRO.2004.838006.
[134] Okada, M., Shinohara, T., Gotoh, T., Ban, S., and Nakamura, Y. “Double Spherical
Joint and Backlash Clutch for Lower Limbs of Humanoids”. In: Proc. IEEE Int. Conf. Robot.
Autom. (ICRA). 2003, pp. 491–6.
[135] Omer, A.M., Ogura, Y., Kondo, H., Morishima, A., Carbone, G., Ceccarelli, M.,
Lim, H., and Takanishi, A. “Development of A Humanoid Robot Having 2-DOF Waist
and 2-DOF Trunk”. In: Proc. IEEE-RAS Int. Conf. Humanoid Robot. (Humanoids). 2005,
pp. 333–8. doi: 10.1109/ICHR.2005.1573589.
[136] Oota, N., Kawasaki, T., and Isozumi, T. “Impact Absorbing Mechanism of Walking
Robot”. Pat. EP 1 550 533 (A1). Kawada Industries, Inc. July 6, 2005.
[137] Ott, Ch., Eiberger, O., Friedl, W., Bäuml, B., et al. “A Humanoid Two-Arm System for
Dexterous Manipulation”. In: Proc. IEEE-RAS Int. Conf. Humanoid Robot. (Humanoids).
2006, pp. 276–83. doi: 10.1109/ICHR.2006.321397.
[138] Pahl, G., Beitz, W., Feldhusen, J., and Grote, K.H. Konstruktionslehre: Grundlagen
erfolgreicher Produktentwicklung. Methoden und Anwendung. 6th ed. (in German). Springer,
2005. isbn: 3-540-22048-8.
[139] Park, I.-W., Kim, J.-Y., Lee, J., and Oh, J.-H. “Mechanical Design of Humanoid Robot
Platform KHR-3 (KAIST Humanoid Robot - 3: HUBO)”. In: Proc. Int. Conf. Humanoid
Robots (Humanoids). 2005, pp. 321–26. doi: 10.1109/ICHR.2005.1573587.
[140] Park, I.-W., Kim, J.-Y., Lee, J., and Oh, J.-H. “Mechanical Design of the Humanoid Robot
Platform, HUBO”. In: Adv. Robotics 21, 11 (2007), pp. 1305–22. doi:
10.1163/156855307781503781.
Bibliography 197
[141] Parker Hannifin Corp. Frameless Motor & Gearmotors. Port Washington, NY USA, 2007.
url: http://www.parkermotion.com/.
[142] Perry, J. Gait Analysis – Normal and Pathological Function. 3rd ed. Slack, 1992. isbn:
1-55642-192-3.
[143] Peterka, R.J. “Sensorimotor Integration in Human Postural Control”. In: J. Neurophysiol.
88, 3 (2002), pp. 1097–1118. url: http://jn.physiology.org/cgi/content/full/88/
3/1097?ijkey=0a502c131800afcfadf15766c60f25289f11994a.
[144] Pfeiffer, F. “The TUM walking machines”. In: Phil. Trans. R. Soc. A 365 (2007),
pp. 109–31. doi: 10.1098/rsta.2006.1922.
[145] Pfeiffer, F. and Cruse, H., eds. Autonomes Laufen. (in German). Berlin, Heidelberg, New
York: Springer, 2004. isbn: 3-540-25044-1.
[146] Pratt, D.J. “Mechanisms of shock attenuation via the lower extremity during running”.
In: Clin. Biomech. 4, 1 (1989), pp. 51–7. doi: 10.1016/0268-0033(89)90068-5.
[147] Pratt, J. and Krupp, B. “Design of a bipedal walking robot”. In: Proc. SPIE. Vol. 6962.
2008. doi: 10.1117/12.777973.
[148] Radcliffe, C.W. “Four-bar linkage prosthetic knee mechanisms: kinematics, alignment
and prescription criteria”. In: Prosthet. Orthot. Int. 18 (1994), pp. 159–73.
[149] Rautscher, G. “Modellierung und Simulation einer geregelten PM-Synchronmaschine
mir Matlab/Simulink”. (in German). Bachelor’s thesis. Lehrstuhl für Angewandte
Mechanik, TU München, 2004.
[150] Regele, R., Levi, P., and Bott, W. “ProRobot – Predicting the Future of Humanoid
Robots”. In: RoboCup 2003: Robot Soccer World Cup VII. Lecture Notes in Computer
Science. Springer, 2004, pp. 366–73. doi: 10.1007/b98623.
[151] Regenstein, K., Kerscher, T., Birkenhofer, C., Asfour, T., Zöllner, M., and
Dillmann, R. “Universal Controller Module (UCoM) – component of a modular concept
in robotic systems”. In: Proc. IEEE Int. Symposium Industrial Electronics (ISIE). 2007,
pp. 2089–94. doi: 10.1109/ISIE.2007.4374930.
[152] Rehbinder, H. and Ghosh, B.K. “Pose estimation using line-based dynamic vision and
inertial sensors”. In: IEEE Trans. Automat. Contr. 48, 2 (2003), pp. 186–99. doi:
10.1109/TAC.2002.808464.
[153] Resetov, D., Pietsch, B., and Gerth, W. “Aufbau des humanoiden Roboters BART III”.
In: Autonome Mobile Systeme 2009. Ed. by Dillmann, R., Beyerer, J., Stiller, C.,
Zöllner, M., and Gindele, T. (in German). Springer, 2009, pp. 209–16. isbn:
978-3-642-10283-7. url:
http://www.springer.com/computer/artificial/book/978-3-642-10283-7.
[154] Rivin, E.I. Mechanical Design of Robots. McGraw-Hill, 1988. isbn: 0-07-052992-2.
[155] Rivin, E.I. Stiffness and Damping in Mechanical Design. Marcel Dekker, 1999. isbn:
0-8247-1722-8.
[156] RoboDrive GmbH. Robodrive: Products. available online:
http://www.robodrive.de/index.php. Seefeld/Germany, Oct. 2008. url:
http://www.robodrive.de/.
[157] Rome, K. “Mechanical properties of the heel pad: current theory and review of the
literature”. In: Foot 8, 4 (1998), pp. 179–85. doi: 10.1016/S0958-2592(98)90026-8.
[158] Rosheim, M.E. Robot Wrist Actuators. John Wiley & Sons, Inc., 1989. isbn: 0-471-61595-1.
[159] R+W Antriebselemente GmbH. Miniatur-Metallbalgkupplungen: Modellreihe MK. (in
German). Klingenberg/Germany, 2007. url: http://www.rw-kupplungen.de/.
[160] R+W Antriebselemente GmbH. Sicherheitskupplungen: Modellreihe SK. (in German).
Klingenberg/Germany, 2007. url: http://www.rw-kupplungen.de/.
198 Bibliography
[161] Sarker, M.O.F., Kim, C.H., Back, S., and You, B.-J. “An IEEE-1394 Based Real-time
Robot Control System for Efficient Controlling of Humanoids”. In: Proc. IEEE/RSJ Int.
Conf. Intell. Robots Syst. (IROS). 2006, pp. 1416–21. doi: 10.1109/IROS.2006.281933.
[162] Saunders, J.B., Inman, V.T., and Eberhart, H.D. “The major determinants in normal
and pathological gait”. In: J. Bone Joint Surg. 35, A (1953), pp. 543–58.
[163] Scheinman, V. and McCarthy, L.M. “Mechanisms and Actuation”. In: Springer
[164] Schmidt, N. “Konstruktion des Fußes für einen humanoiden Roboter”. (in German). Term
project. Lehrstuhl für Angewandte Mechanik, TU München, 2006.
[165] Schröder, D. Elektrische Antriebe – Grundlagen. 3rd ed. (in German). Berlin, Heidelberg,
New York: Springer, 2007. isbn: 978-3-540-72764-4.
[166] Schröder, D. Elektrische Antriebe – Regelung von Antriebssystemen. 2nd ed. (in German).
Berlin, Heidelberg, New York: Springer, 2001. isbn: 3-540-41994-2.
[167] Schwienbacher, M. “Entwicklung eines Kraft-Momentensensors für einen Humanoiden
Roboter”. (in German). Diploma thesis. Lehrstuhl für Angewandte Mechanik, TU
München, 2007.
[168] Schwienbacher, M. “Entwicklung und Konstruktion des Beckens für einen humanoiden
Roboter mit Hilfe der Topologieoptimierung”. (in German). Term project. Lehrstuhl für
Angewandte Mechanik, TU München, 2006.
[169] Seara, J.F., Schmidt, G., and Lorch, O. “ViGWaM Active Vision System – Gaze Control
for Goal-Oriented Walking”. In: Proc. Int. Conf. Climbing and Walking Robots (CLAWAR).
2001, pp. 265–72.
[170] Sellaouti, R., Stasse, O., Kajita, S., Yokoi, K., and Kheddar, A. “Faster and Smoother
Walking of Humanoid HRP-2 with Passive Toe Joints”. In: Proc. IEEE/RSJ Int. Conf. Intell.
Robots Syst. (IROS). 2006, pp. 4909–14. doi: 10.1109/IROS.2006.282449.
[171] Shimano, B. and Roth, B. “On force sensing information and its use in controlling
manipulators”. In: Proc. IFAC Int. Symp. Information-Control Problems. 1977, pp. 119–26.
[172] Siciliano, B. and Khatib, O., eds. Springer Handbook of Robotics. Berlin, Heidelberg, New
York: Springer, 2008. isbn: 978-3-540-23957-4.
[173] SKF Linearsysteme GmbH. Ball screws. Product catalog. Schweinfurt/Germany, June
2008. url: http://www.linearmotion.skf.com/.
[174] Slatter, R., ed. Leichtbau in der Antriebstechnik. (in German). Aachen: Shaker, 2004. isbn:
3-8322-2805-5.
[175] Sodeyama, Y., Nishino, T., Namiki, Y., Nakanishi, Y., Mizuuchi, I., and Inaba, M.
“The designs and motions of a shoulder structure with a spherical thorax, scapulas and
collarbones for humanoid ‘Kojiro’”. In: Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS).
2008, pp. 1465–70. doi: 10.1109/IROS.2008.4651221.
[176] Sugahara, Y., Hashimoto, K., Sunazuka, H., Kawase, M., Ohta, A., Tanaka, C.,
Lim, H.-o., and Takanishi, A. “WL-16RII: Prototype of Biped Walking Wheelchair”. In:
Proc. CISM-IFToMM Symposium on Robot Design, Dynamics, and Control (RoManSy). 2006,
pp. 313–20.
[177] Taghirad, H.D. and Bélanger, P.R. “Modeling and parameter identification of harmonic
drive systems”. In: J. Dyn. Sys., Meas., Control 120, 4 (1998), pp. 439–44. doi:
10.1115/1.2801484.
[178] Tajima, R., Honda, D., and Suga, K. “Fast Running Experiments Involving a Humanoid
Robot”. In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 2009, pp. 1571–6.
Bibliography 199
[179] Takahashi, M., Nishiwaki, K., Kagami, S., and Mizoguchi, H. “Development and
Evaluation of An Attitude Measuring System that uses Acceleration Information of
Walking for Biped Robots”. In: Proc. IEEE Conf. on Sensors (Sensors). 2007, pp. 292–5. doi:
10.1109/ICSENS.2007.4388394.
[180] Takao, S., Ohta, H., Yokokohji, Y., and Yoshikawa, T. “Function analysis of human-like
mechanical foot, using mechanically constrained shoes”. In: Proc. IEEE/RSJ Int. Conf. Intell.
Robots Syst. (IROS). 2004, pp. 3847–52. doi: 10.1109/IROS.2004.1390014.
[181] Takenaka, T., Gomi, H., Hamaya, K., Akimoto, K., Shirokura, S., Asatani, M., and
Tanaka, K. “Leg Joint Assist Device of Legged Mobile Robot”. Pat. EP 1 820 612 (A1).
Honda Motor Co., Ltd. Aug. 22, 2007.
[182] Takenaka, T., Gomi, H., S., Shigemi, and Matsumoto, T. “Leg-type Moving Robot and
Floor Reaction Force Detection Device Therefor”. Pat. EP 1 475 193 (A1). Honda Motor
Co., Ltd. Nov. 10, 2004.
[183] Takenaka, T., Matsumoto, K., Gomi, H., and Hamaya, K. “Leg Type Moving Robot”.
Pat. EP 1 702 725 (A1). Honda Motor Co., Ltd. Sept. 20, 2006.
[184] Takenaka, T., Matsumoto, K., Gomi, H., and Hamaya, K. “Legged Mobile Robot”. Pat.
EP 1 785 235 (A1). Honda Motor Co., Ltd. May 16, 2007.
[185] Takenaka, T., Matsumoto, T., Yoshiike, T., and Shirokura, S. “Real time motion
generation and control for biped robot —2nd report: Running gait pattern generation—”.
In: Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS). 2009, pp. 1092–9. doi:
10.1109/IROS.2009.5354654.
[186] Tellez, R., Ferro, F., Garcia, S., Gomez, E., Mora, J.D., Pinyol, D., Poyatos, J.,
Torres, O., Velazquez, J., and Faconti, D. “Reem-B: an Autonomous Lightweight
Human-size Humanoid Robot”. In: Proc. IEEE-RAS Int. Conf. Humanoid Robot.
[187] Tränkler, H.-R. Sensortechnik. Handbuch für Praxis und Wissenschaft. Ed. by
Obermeier, E. Springer, 1998. isbn: 978-3-540-58640-1.
[188] Tsagarakis, N.G., Metta, G., Sandini, G., Vernon, D., et al. “iCub: The Design and
Realization of an Open Humanoid Platform for Cognitive and Neuroscience Research”. In:
Adv. Robotics 21, 10 (2007), pp. 1151–75. doi: 10.1163/156855307781389419.
[189] Uchiyama, M., Bayo, E., and Palma-Villalon, E. “A Systematic Design Procedure to
Minimize a Performance Index for Robot Force Sensors”. In: J. Dyn. Sys., Meas., Control
113, 3 (1991), pp. 388–94. doi: 10.1115/1.2896422.
[190] Umberger, B.R. “Effects of suppressing arm swing on kinematics, kinetics, and energetics
of human walking”. In: J. Biomech. 41, 11 (2008), pp. 2575–80. doi:
10.1016/j.jbiomech.2008.05.024.
[191] Vanderborght, B., Verrelst, B., Ham, R. van, Damme, M. van, Lefeber, D., Meira
Y Duran, B., and P., Beyl. “Exploiting Natural Dynamics to Reduce Energy Consumption
by Controlling the Compliance of Soft Actuators”. In: Int. J. Robot. Res. 25, 4 (2006),
pp. 343–58. doi: 10.1177/0278364906064566.
[192] VDI 2206. Design methodology for mechatronic systems. June 2004.
[193] VDI 2209. 3D-product modeling. June 2006.
[194] Vishay Micro-Measurements, Inc. Fatigue Characteristics of Vishay
Micro-Measurements Strain Gages. Tech Note. Malvern, PA (USA), 2007. url:
http://www.vishay.com/doc?11058.
[195] Voyles, R., Morrow, J., and Khosla, P. “The Shape from Motion Approach to Rapid and
Precise Force/Torque Sensor Calibration”. In: J. Dyn. Sys., Meas., Control 119, 2 (1997),
pp. 229–35. doi: 10.1115/1.2801238.
200 Bibliography
[196] Vucobratović, J. and Borovac, B. “Zero-Moment Point – Thirty Five Years Of Its Life”.
In: Int. J. Humanoid Robot. 1, 1 (2004), pp. 157–73. doi: 10.1142/S0219843604000083.
[197] Vukosavić, S.N. and Stojić, M.R. “Suppression of torsional oscillations in a
high-performance speed servo drive”. In: IEEE Trans. Ind. Electron. 45, 1 (1998),
pp. 108–17. doi: 10.1109/41.661311.
[198] Watson, P.C. and Drake, S.H. “Pedestal and wrist force sensors for automatic assembly”.
In: Proc. Int. Symp. Industrial Robots. 1975, pp. 501–11.
[199] Whittle, M.W. “Generation and attenuation of transient impulsive forces beneath the
foot: a review”. In: Gait Posture 10, 3 (1999), pp. 264–75. doi:
10.1016/S0966-6362(99)00041-7.
[200] Whittle, M.W. and Levine, D. “Three-dimensional relationships between the
movements of the pelvis and lumbar spine during normal gait”. In: Hum. Mov. Sci. 18, 5
(1999), pp. 681–92. doi: 10.1016/S0167-9457(99)00032-9.
[201] Wisse, M., Schwab, L., and Van der Helm, F.L.T. “Passive walking dynamic model with
upper body”. In: Robotica 22, 6 (2004), pp. 681–8. doi: 10.1017/S0263574704000475.
[202] Wittenstein cyber motor GmbH. Full-line catalog. Igersheim/Germany, 2003. url:
http://www.wittenstein-cyber-motor.de/.
[203] Wright, D.G., Desai, S.M., and Henderson, W.H. “Action of the subtalar and
ankle-joint complex during the stance phase of walking”. In: J. Bone Joint Surg. Am. 46, A
(1964), pp. 361–82.
[204] Yamaguchi, J., Soga, E., Inoue, S., and Takanishi, A. “Development of a Bipedal
Humanoid Robot–Control Method of Whole Body Cooperative Dynamic Biped Walking”.
In: Proc. IEEE Int. Conf. Robot. Autom. (ICRA). 1999, pp. 368–74. doi:
10.1109/ROBOT.1999.770006.
[205] Yamaguchi, J., Takanishi, A., and Kato, I. “Experimental Development of a Foot
Mechanism with Shock Absorbing Material for Acquisition of Landing Surface Position
Information and Stabilization of Dynamic Biped Walking”. In: Proc. IEEE Int. Conf. Robot.
[206] Yamamoto, K., Sugihara, T., and Nakamura, Y. “Toe Joint Mechanism Using Parallel
Four-bar Linkage Enabling Humanlike Multiple Support at Toe Pad and Toe Tip”. In: Proc.
IEEE-RAS Int. Conf. Humanoid Robot. (Humanoids). 2007.
[207] Yamazaki, Y., Kobayashi, N., and Koshiishi, T. “Head Support Structure of Humanoid
Robot”. Pat. JP 2006-167 832 (A) (Japanese). Honda Motor Co., Ltd. June 29, 2006.
[208] Yoon, J., Nandha, H., Lee, D., and Kim, G. “A Novel 4-DOF Robotic Foot Mechanism
with Multi-platforms for Humanoid Robot (SICE-ICCAS 2006)”. In: Proc. Int. Joint Conf.
SICE-ICCAS. 2006, pp. 3500–4. doi: 10.1109/SICE.2006.314621.
[209] Zhang, G. and Furusho, J. “Speed control of two-inertia system by PI/PID control”. In:
IEEE Trans. Ind. Electron. 47, 3 (2000), pp. 603–9. doi: 10.1109/41.847901.

PDF

Uploaded by

Copyright:

Available Formats

PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PDF

Uploaded by

Copyright:

Available Formats

TECHNISCHE UNIVERSITÄT MÜNCHEN

Lehrstuhl für Angewandte Mechanik

Design and Realization of a

Dipl.-Ing. Univ. Sebastian Lohmeier

Vollständiger Abdruck der von der Fakultät für Maschinenwesen der

Vorsitzender: Univ.-Prof. Dr.-Ing. Udo Lindemann

Prüfer der Dissertation:

1. Univ.-Prof. Dr.-Ing. habil. Heinz Ulbrich

2. Univ.-Prof. Dr.-Ing. Horst Baier

© 2010 by Sebastian Lohmeier. All rights reserved

Author contact: [email protected]

This thesis is also available as a printed book published by

Keywords: humanoid robots, anthropomorphic robots, legged locomotion, mecha-

Stichworte: humanoide Roboter, anthropomorphe Roboter, zweibeiniges Laufen,

during the countless public performances of Johnnie in 2003/04. The designer of

2 Mechatronic Design Concept 17

2.7.3 External Forces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

4 Sensor System 133

5 Electronics Architecture 159

5.3 Communication System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

6 Lessons Learned 167

7 Conclusions and Future Work 171

A Anatomic and Biomechanical Fundamentals of Human Locomotion 175

B Main Dimensions 177

C Technical Data of the Joint Actuators 179

D Load Profiles 181

No matter whether one believes in God’s creation or theory of evolution: everyone

1.1 Literature Review

© Atsuo Takanishi Lab., Waseda University

© Toyota Motor Corporation

(a) Wabian-2R (b) Asimo (c) Toyota running robot

Figure 1.1: Humanoid robots, part 1

1.1.1 Fully Actuated Biped Walking Robots

© JSK Robotics lab., The University of Tokyo

© KAWADA Industries, Inc. / AIST

(a) HRP-2 (b) HRP-3 (c) H7

Figure 1.2: Humanoid robots, part 2

© Institute of Applied Mechanics, TU München

© JST-ATR Computational Neuroscience Lab.

(a) Kojiro (b) Johnnie (c) CB-i

Figure 1.3: Humanoid robots, part 3

complex structure is indeed challenging, both in building and control.

commands in case the vision system is not connected.

by controlling the compression of the springs which is measured by quadrature

1.1.2 Balance Control of Fully Actuated Biped Robots

1.1.3 Passive Dynamic Walkers

of the elastic elements are identified by numerical optimization using different

1.1.4 Other Relevant Developments

Design and Realization

1.2 Contributions of this Thesis

development relies on the “classical” approach, using well-proven and state-of-the-

Humanoid robots are prime examples of mechatronic systems. They consist of a

2.1 Systematic Approach to Managing Complexity

2.1.1 Methodology in the Development Process

Customer Needs Product Architecture

Specifications Detailed Layout

2.1.2 Project Planning

and/or chronological dependencies among work packages are identified, and

2.2 Kinematic Structure

ϕ̇ PRS ϕ̇ PRS 2π l˙PRS