658 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 14, NO.

6, DECEMBER 2009

System Design and Control of Anthropomorphic

Walking Robot LOLA
Sebastian Lohmeier, Thomas Buschmann, and Heinz Ulbrich

Abstract—This paper presents the 25-DOF full-size humanoid

robot LOLA. Our goal is to realize fast and human-like walking.
Furthermore, we want to increase the robot’s autonomous, vision-
guided walking capabilities. LOLA is characterized by a redundant
kinematic configuration, an extremely lightweight design, joint ac-
tuators with brushless motors and an electronics architecture using
decentralized joint control. Special emphasis was put on an im-
proved mass distribution to achieve good dynamic performance.
Center of mass trajectories are calculated in real-time from foot-
step locations using a spline collocation method. Reference trajec-
tories are modified by a stabilizing control system based on hybrid
force/position control with an inner joint position control loop.
Index Terms—Anthropomorphic robots, humanoid robots,
legged locomotion, robot design, robot dynamics, walking control.

I. INTRODUCTION
CCORDING to a recent paper on the robotics market,
A humanoid service robots might be an emerging applica-
tion area [1]. The anthropomorphic structure is one reason why
such systems are ideal general-purpose assistant robots for typ-
ical everyday tasks. Other scenarios include the entertainment
industry and academic education and research. Fig. 1. Photograph of the 25-DOF humanoid walking robot LOLA.
Quite a few sophisticated full-size humanoid robots are devel-
oped by companies, like Toyota’s running robot [2] and Honda’s TABLE I
ASIMO [3], while most are developed by research laboratories HARDWARE SPECIFICATION
and universities, for example, H7 [4], HRP-2 [5] and HRP-3 [6]
(AIST and Kawada Industries), WABIAN-2 [7], HUBO [8], and
JOHNNIE [9].
For everyday tasks in production, office or home environ-
ments, stable and fast biped locomotion is a basic skill. Com-
pared with human beings, higher walking speeds, and flexible
motion generation still remain challenging due to many un-
solved control problems, e.g., fast walking and running [2], [10],
sudden turning motions, walking on rough terrain and trajectory
generation in complex environments. On the other hand, the
robot hardware contributes significantly to system performance
components of the mechanical design are presented. The second
if designed thoroughly: both robot hardware and software must
part deals with the robot simulation, and outlines the real-time
be seen as tightly coupled parts of a highly integrated mecha-
trajectory generation and stabilizing control.
tronic system.
The first part of this paper gives an overview of the mecha-
tronic system: the fundamental design considerations and key II. SYSTEM OVERVIEW
LOLA is 180 cm tall and weighs approximately 55 kg. The
Manuscript received February 13, 2009; revised July 1, 2009. First pub- physical dimensions of the robot are based on anthropometric
lished October 9, 2009; current version published November 11, 2009. Recom- data. Fig. 1 shows a photograph of the robot, the key data are
mended by Guest Editor M. Goldfarb. This work was supported by the Deutsche summarized in Table I. The distinguishing characteristics of
Forschungsgemeinschaft under Grant UL 105/29.
The authors are with the Institute of Applied Mechanics, Technical Univer- LOLA are the redundant kinematic structure with 7-DOF legs, an
sity Munich, 85748 Garching, Germany (e-mail: [email protected]; extremely lightweight construction, and a modular joint design
[email protected]; [email protected]). with high power density based on brushless motors. The mass
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. distribution of the legs is improved to achieve good dynamic
Digital Object Identifier 10.1109/TMECH.2009.2032079 performance.
1083-4435/$26.00 © 2009 IEEE
LOHMEIER et al.: SYSTEM DESIGN AND CONTROL OF ANTHROPOMORPHIC WALKING ROBOT LOLA 659

2) Force/Torque Sensors: Two custom-made six-axis force/

torque sensors are tightly integrated into the foot structure. The
sensor body is a monolithic aluminum component with four
shear beams in a Maltese cross arrangement. Each beam holds
two pairs of strain gauges connected to half bridges. Thin mem-
branes decouple the beam deflections to a far extent and reduce
crosstalk. In order to protect the sensor from damage during ex-
periments, mechanical end stops are integrated as an overload
protection. A calibration error of less than 0.5% is achieved by
applying many different load cases. At a total weight of 395 g
the sensor includes an overload protection and all necessary
electronics.
Fig. 2. Kinematic structure of LOLA. 3) Inertial Measurement Unit (IMU): The IMU estimates
the orientation and angular velocities of the upper body. Since
accuracy and signal quality of the IMU considerably affect the
A. Kinematic Structure performance of the stabilizing controller, we have chosen a
Simulations and experiments have shown that additional, re- commercial high-precision IMU (iVRU-FC-C167, iMAR Nav-
dundant DOFs enable more natural and flexible gait patterns igation) with fiber-optic gyroscopes and MEMS accelerometers.
and extend the abilities of the robot in general. Fig. 2 shows the
kinematic configuration with 25 actuated DOFs: the legs have III. MECHATRONIC DESIGN CONCEPT
7 DOFs each, the pelvis has 2 and each arm has 3 DOFs. The
3-DOF stereo camera head consists of a pan/tilt unit. In addi- LOLA’s hardware approach is based on experiments with
tion, the angle of convergence can be adjusted, enabling stereo JOHNNIE and tries to settle most of the technical problems dis-
vision of objects close to the robot. covered during a thorough hardware analysis. The fundamental
Considering the leg and pelvis joints, the kinematic chain of design concept is inspired by research on human walking and
the upper body with respect to the stance foot has a redundant the underlying mechanisms. State-of-the-art mechatronics tech-
structure with 9 DOFs. Kinematic redundancies augment the nology is used to create a highly dynamic robot with enhanced
robot’s agility, but they also contribute to reducing joint loads. walking capabilities and autonomy.
LOLA has active toe joints that allow to control the center of
pressure during heel rise, whereas robots with monolithic feet A. Design Process
or passive toe joints would be marginally stable. Area contact Fast locomotion poses a significant challenge for biped robots
of the toe segment stabilizes the robot and facilitates forward and requires a thorough design of the mechatronic system. The
roll across the forefoot. Moreover, heel lift-off during termi- robot development follows the procedure described in guide-
nal stance, which occurs shortly before the swing leg contacts line VDI 2206 “Design methodology for mechatronic systems”
the ground, can also reduce joint loads. As found by Kerrigan [13]. It is an iterative and open-ended process of design and
et al. [11], heel rise in human gait contributes significantly to simulation.
reducing the center of mass (CoM) vertical displacement. Due After certain design milestones have been completed, new
to the importance of stance leg heel rise in human walking, the inertia properties of the links and actuators are obtained from
idea of implementing toe joints on a humanoid robot is not new. the 3-D-computer-aided design (CAD) model. These are used
Yet, there are very few humanoid robots with actively driven toe to calculate joint loads, workspaces and constraint forces us-
joints, e.g., [2], [12]. ing the dynamics simulation of the robot (Section VI), which
are the basis for the dimensioning of actuators and structural
B. Sensor System components. In this way, both simulation and CAD model are
iteratively refined.
The sensor system supports the implementation of model-
based control algorithms and is optimized for signal quality and
B. Special Requirements for High-Speed Walking
bandwidth.
1) Joint Sensors: Absolute angular sensors on the output Without doubt, many open questions make the control of fast
shafts of all joints compensate elasticities and nonlinearities walking very challenging. Increased robustness and flexibility
in the drivetrains and enable the robot to (theoretically) start are key components to guarantee stability of the robot in differ-
from arbitrary positions. We use Heidenhain EnDat interface ent situations. This is achieved by an integrated control approach
(ECI) encoders with a resolution of 17 bit and an accuracy of which combines the stabilization of global dynamics by manipu-
0.1◦ , which are used as a resolver replacement in industrial lation of contact forces and hybrid force/position control in task
motor control. The cutoff frequency is 6 kHz and the latency for space (Section VIII). Fast walking also requires an improved
continuous sampling of position values is 5 µs. Thus, the delay real-time walking pattern generation. A new method is used to
in position feedback is minimal and the velocity measurement calculate the reference trajectories for the following three steps
can be done by numerical differentiation. (Section VII).
660 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 14, NO. 6, DECEMBER 2009

Moreover, several aspects in the design of the mechanical sys- TABLE II

GEAR SIZES, REDUCTION RATIOS AND JOINT WORKING RANGES
tem can improve system performance. Besides a revised kine-
matic structure, three design objectives are defined to improve
leg dynamics: 1) sufficient mechanical stiffness; 2) high CoM;
and 3) low moments of inertia of the leg links.

C. Design Features to Improve System Dynamics

1) Mechanical Stiffness: Both static and dynamic behaviors
of the mechanical system have a strong influence on the walking
capabilities. In particular, structural weak points of the locomo-
tor system can degrade walking performance by causing struc-
tural vibrations and deviations from the reference trajectories,
which can destabilize the robot.
The basic design objective is, therefore, to balance structural
stiffness and actuator performance with lightness of the mechan-
ical parts. To avoid a particular range of resonant frequencies,
these considerations cannot be limited to the component-level.
Rather, the elastodynamic dimensioning of the robot structure
must include the drivetrains and other components within the
load path, such as bolted flange connections and link bearings. dynamic performance of the system in general; 2) the accuracy
For instance, stiffness-based gear selection is carried out in ad- of model simplifications; and 3) the performance of stabilizing
dition to the “standard” procedure of torque-speed-based di- control.
mensioning in order to minimize oscillations of the feet. With Our approach is to design major structural components as
the torsional stiffness K of the gear and an effective moment of investment castings made from aluminum. In order to meet the
inertia J, the resonance frequency at the gear output is weight and stiffness targets, design proposals are created by

topology optimization [15]. Moreover, actuator performance is
1 K increased carefully (Section IV-A) and new kinematic struc-
f= . (1) tures are developed for the knee (Section IV-B) and ankle joints
2π J
(Section IV-C), so that heavier component parts can be placed
As an empirical value recommended by the gear manufacturer,
close to the hip joint axis.
f should be above 15 Hz.
2) CoM Height: Unlike humans, the largest portion of a
biped robot’s weight resides in its legs, since motors and gears IV. MECHANICAL DESIGN
determine approximately a third of the overall weight. The CoM A. Actuator Design
height is therefore lower than in humans, typically at the height
of the hip joint or below. According to the linear inverted pen- The joint actuators account for about a third of the total
dulum mode by Kajita et al. [14], one possible solution for CoM weight, making the development of compact and lightweight
lateral motion are pieces of hyperbolic curves actuators crucial. We use high-performance permanent magnet

 synchronous motors (PMSM) because of their superior torque
g and speed capabilities. These motors allow to either increase
∆yCoM ∝ cosh Ts . (2)
zCoM actuator performance without adding additional weight, or de-
crease actuator mass without decreasing performance. The joints
According to this model, the CoM lateral swing ∆yCoM de- are driven through harmonic drive gears and planetary roller
creases with higher CoM positions zCoM for a given single sup- screws, respectively. Table II shows the gear sizes, reduction
port period Ts . Especially at higher walking speeds, the stability ratios, and working ranges of all joints. Further details on the
of the robot is increased if ∆yCoM is small, because of reduced actuator design have been published in [16].
oscillations of the angular momentum around the longitudinal
axis. A more human-like mass distribution could be achieved by
adding weight to the upper body, which would in turn require B. Knee Joint
larger actuators and increase the weight of the leg segments. A harmonic drive-based actuator located in the knee joint
Our approach is therefore to shift the total CoM as close as pos- axis would unacceptably increase the thigh moment of inertia,
sible to the hip joint while keeping the robot’s overall weight and a large part of the enhanced hip joint output would be
minimal. spent on accelerating the heavier thigh. By employing a roller
3) Leg Inertia: Mass distribution in the leg apparatus is a screw-based linear drive as shown in Fig. 3, mass distribution
design variable which not only affects CoM position, rather, the in the hip-thigh area is significantly improved: the motor is
influence on the inertia of individual leg segments is much more located close to the hip joint axis, reducing the thigh inertia
significant. Minimizing the resulting leg inertia and reducing its by 65%, and the mass of the actuator itself by more than 10%,
posture dependence is an important way of improving: 1) the without reducing performance. The four-bar linkage mechanism
LOHMEIER et al.: SYSTEM DESIGN AND CONTROL OF ANTHROPOMORPHIC WALKING ROBOT LOLA 661

Fig. 3. Mechanism employed for the knee joint (left). Torque and speed re-
quirements of knee joint, human torque capacity adapted from [17] (right).

is backlash-free, back-drivable and has a nonlinear torque-speed

characteristic similar to the human knee (Fig. 3, right-hand side):
the maximum torque occurs around 55◦ , which is advantageous
for typical gait patterns. Conversely, maximum speeds increase
at a stretched leg configuration where they are needed.
Fig. 4. Ankle joint actuation by a 2-DOF parallel mechanism.
C. Ankle Joint
The ankle joint axes differ clearly in their torque-speed char-
segments of the leg kinematic chains and undergo the largest
acteristics. By employing parallel drives, the motor peak torques
accelerations. The mechanical structure must, therefore, be ex-
can be reduced by about 35% [15]. Fig. 4 explains the function-
tremely lightweight, yet it must be able to withstand shock loads
ality of the ankle joint actuation: the ankle joint (3) is actuated by
of a multiple of the robot’s weight during initial contact.
two spatial slider-crank mechanisms (7)–(9) with 1 DOF each,
Despite the obvious fact that foot dynamics have a strong
where the motors (4) are mounted on the thigh (1), close to the
influence on the walking performance, there is surprisingly little
hip joint. Compared to a previous design concept [15], the joint
detailed literature. Most robotic feet consist of a monolithic
workspace is increased and shank inertia with respect to the hip
baseplate covered with rubber or plastic to increase friction
joint is reduced by more than 20%.
and use compression/shear mounts as shock absorbers [3], [5].
Each linear drive is composed of a roller screw (8) mounted
Under shear loading, however, the stiffness of such elements
on the shank and a linear bearing (9) to keep the screw free from
is considerably lower than under pressure loading which can
radial loads. The link (10) connects the linear carriage (7) and
easily destabilize the robot during stance.
the foot segment. The synchronous belt (5) connects the motor
Developing a functional foot design is not trivial, thus, it is
shaft (4) to the input shaft of the bevel gear (6) in the knee
useful to take the main functions of the human foot into account.
joint axis. The output shaft of (6) finally drives the roller screw
The main requirements are:
(8), which is arranged in longitudinal direction of the shank
1) high grip on different surfaces, to ensure proper ground
and perpendicular to the knee joint axis. The absolute angular
contact;
sensors (11) allow direct measurement and control of the joint
2) passively compliant elements with a) good damping char-
angles.
acteristics for shock attenuation at initial contact and b)
The planetary roller screws are preloaded and free from back-
progressive spring characteristics at ground contact to en-
lash. The whole mechanism is back-drivable.
sure effectiveness of stabilizing control;
The longer transmission distance reduces inertia and gravity
3) rigid mechanical structure to transmit propulsive forces at
loading. On the other hand, the spatial separation of electric
a bandwidth sufficient for system stabilization;
motor and transmission requires the use of auxiliary transmis-
4) compensate smaller unevenness;
sions and couplings, which may introduce elasticity and reduce
5) minimal weight.
system stiffness. To achieve high transmission accuracy without
From these requirements, the design of the robotic foot shown
adding too much elasticity and backlash to the drivetrain, a syn-
in Fig. 5 is derived. The toe joint (1) divides the foot into forefoot
chronous belt (5) with a high tensile stregth and an optimized
(2) and hallux (3). The force/torque sensor (4) is designed as an
tooth profile is chosen. The bevel gear (6) is a commercial pre-
integral, load-bearing element within the forefoot. The modular
cision gear with minimal backlash (8 arc min), optimized for
mechanical configuration with the bolted flange connections (5)
position servo applications.
facilitates experiments with different foot geometries and allows
to easily adjust the support polygon.
D. Foot Dynamics
Several mechanisms are implemented for shock attenuation:
During normal locomotion, the feet are the only parts of the viscoelastic layers (6) with good damping characteristics are
robot which are subject to external loads. They are the last used in analogy to the fat pads in the human heel and toe for
662 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 14, NO. 6, DECEMBER 2009

Fig. 6. Electronics architecture of LOLA.

mation is forwarded to the CCU. Local controllers and CCU are

connected via the Ethernet-based real-time communication sys-
tem Sercos-III. The CCU is based on a PC platform (Core 2 Duo
Fig. 5. Mechanical design of the feet: besides viscoelastic members, the pas- Mobile, 2.33 GHz), running the QNX real-time operating sys-
sive heel segment with hydraulic damper contributes to shock attenuation. tem. The local controllers are a custom development because of
compactness and various sensor–actuator interfaces.
Gait generation and stabilization run on the on-board com-
damping the impact transients and compensating small uneven-
puter system without any support from outside except for power
ness. The foot sole (7) is made of wear-resistant, high-friction
supply. An external PC is used only for monitoring purposes and
rubber. For improved energy dissipation during initial contact,
to give basic operating commands if the robot is not connected to
a passive heel segment (8) similar to the human hindfoot and
the vision system. Due to high computing power requirements,
a hydraulic shock absorber (10) are implemented. The four-
vision processing is done on an external PC cluster.
bar linkage (11) adjusts the strokes of heel and damper. The
heel bearing (9) corresponds to the subtalar joint of the human
foot [18]. The mechanical end stop (11) ensures the proper trans- VI. DYNAMICS SIMULATION
mission of propulsive forces and stabilizing moments while in A key tool for hardware and controller design is a dynam-
ground contact. The feasibility of the foot concept has been ics simulation for calculating loads, analyzing system dynam-
proved in preliminary tests with JOHNNIE. ics, control system robustness, and performance. This section
The functionality of the foot in the different gait phases is presents a refined version of the simulation system described
illustrated in Fig. 5 (left): At initial contact (a) the heel segment in [19], which is used for developing and testing LOLA’s hard-
protrudes from the foot sole. Its passive motion increases the ware and control algorithms. We model the robot as three cou-
time over which the foot is decelerated and during which en- pled dynamical systems: 1) rigid multibody dynamics; 2) contact
ergy is dissipated by the hydraulic damper. At loading response dynamics; and 3) drive dynamics.
(b) the heel segment is retracted and more impact energy is
dissipated by the deformation of the viscoelastic layer. Ground A. Rigid Multibody Dynamics
contact during midstance (c) is ensured by both heel and hallux,
while the heel is lifted in terminal stance (d). The equations of motion (EoM) for the multibody sys-
As in humans, passive and active mechanisms of shock ab- tem (MBS) are calculated recursively using the Newton–Euler
sorption complement each other: the compliant contact elements method, exploiting the robot’s tree structure to increase effi-
and the heel damper act as a mechanical low-pass for the trans- ciency. The result is a set of nonlinear ordinary differential
ferred forces. Although it is necessary to attenuate impact tran- equations (ODEs) describing the rigid multibody dynamics
sients, compliance in the foot–ground contact bounds the me-
M q̈ + h = W λλ + W τ τ . (3)
chanical bandwidth of transferable forces and moments. Passive
shock absorption is therefore limited by the bandwidth required Here, M is the mass matrix and h the vector of smooth forces.
for stabilizing the upper body orientation and is complemented λ are the contact forces acting on the MBS via the Jacobian
by active mechanisms, cf., Section VIII. W λ , while τ and W τ are the actuator forces and actuator force
Jacobian, respectively.
V. ELECTRONICS ARCHITECTURE
Fig. 6 gives a schematic overview of the electronics archi- B. Contact Dynamics
tecture. LOLA is controlled by a central control unit (CCU) The robot contacts the ground with its feet which are equipped
mounted on the upper body and nine local controllers carrying with compliant, shock absorbing materials (cf., Section IV-D).
out low-level tasks, such as joint control and sensor data pro- During walking the elastic contact elements are compressed
cessing. Similar to hierarchical structures in biological systems, by several millimeters, greatly influencing system dynamics. In
sensor data is preprocessed decentrally and only relevant infor- simulation, these materials are modeled as linearly viscoelastic
LOHMEIER et al.: SYSTEM DESIGN AND CONTROL OF ANTHROPOMORPHIC WALKING ROBOT LOLA 663

gN ,i = proxC N (fN ,i − rġN ,i ) ∀i ∈ Ic

f T ,i = proxC T (f T ,i − rġ T ,i ) ∀i ∈ Ic (8)

where Ic = {i|gN ,i = 0} is the set of active contacts. The non-

linear complementarity conditions (5) and (6) were rewritten
using proximal point functions [20].
The unilateral contact acts on the robot via the massless con-
tact elements. The states therefore are strictly continuous, but
their time derivatives are discontinuous making the robot is a
Fillipov-type system.
We solve (4), (5), and (6) for f , d˙ using a fixed point itera-
Fig. 7. Model of compliant ground contact.
tion scheme. Because of the discontinuity, we use a first-order
method with fixed step size to calculate the time evolution of
with negligible mass, i.e., their dynamics are governed by (8).
B d˙ + Kd = f (4)
VII. TRAJECTORY GENERATION
where K and B are stiffness and damping matrices, d is node
displacements, and f is node forces. For diagonal K and B, The real-time walking pattern generator calculates a trajectory
(4) describes decoupled point contacts, while the general case for the following three steps at the beginning of each step. To
describes a linear finite-element model. The contact elements facilitate real-time execution, we separate planning into smaller
are rigidly attached to one body of the MBS, either a heel or and simpler subtasks. First, constraints are calculated from given
a toe segment. Forces at the nodes attached to the MBS are parameters such as average walking speed. Second, foot trajec-
summed up to the subvector λi of the full contact force vector tories are planned in task space. We then calculate reference
λT = (λ λT0 , . . . , λTM −1 ) (Fig. 7). torques that minimize Ṫi,ref 2
dt, i ∈ x, y and stay within the
Free nodes may touch the ground, but not penetrate it, i.e., admissible range. By choosing a piecewise linear parametriza-
there is a unilateral contact between free nodes and environment tion of Ti,ref , we obtain a quadratic programming problem with
bound constraints, that can be solved efficiently. Finally, the
fN ,i ≥ 0 ∧ gN ,i ≥ 0 ∧ fN ,i gN ,i = 0 (5) robot’s CoM trajectory is calculated based on the previous cal-
fN ,i − µf T ,i  ≥ 0 ∧ (fN ,i − µf T ,i )ġ T ,i  = 0. (6) culation steps [21].

Here, µ is the coefficient of friction, fN ,i and f T ,i are force

A. Robot Model
components normal to the contact plane and in the contact plane,
respectively, while gN ,i is the minimum distance between the For real-time planning, the robot is modeled as a system of
ith node and the environment, and ġ T ,i is the tangential speed three point masses: one for each leg and a third for the body. This
of the node relative to the environment. is a modification of the well-known inverted pendulum model
(InvPM). Additional point masses to model swing leg dynamics
C. Drive Dynamics have been proposed in, e.g., [22], [23].
The simplified EoM in the lateral y–z-plane is given by
We assume a field-oriented control of the PMSM, where
the magnetizing currents I d = 0 are controlled ideally and the mb [zb ÿb − yb (z̈b + g)]
torque producing currents I q = I are used for torque control. 
This allows a simplified modeling of the PMSM as dc motors = −Tx + ml [yl,i (z̈l,i + g) − zl,i ÿl,i ] (9)
i=1..2
LI˙ + RI + kM ω rot = U . (7)
where b denotes the body mass point, i is the foot index, and ml
Here, U is the armature voltage, ω rot the rotor’s angular ve- is the mass of one foot mass point. In the following, we will de-
locity, and L, R, and kM are the inductance, resistance, and scribe the trajectory planning procedure for the lateral direction
motor back electromotive force (EMF) constant, respectively. only. The procedure for the sagittal direction is identical.
The actuator force is then given by τ = kM I + τ gear , where Since the swing leg trajectory is calculated independently
τ gear is the friction force created by reduction gears [19]. from the CoM trajectory, (9) can be written in the form of
the InvPM by defining a pseudo contact torque T̃x := Tx −
D. Time Integration ml i=1..2 [yl,i (z̈l,i + g) − zl,i ÿl,i ].
Putting all elements together, we obtain the following set of
equations describing the biped robot dynamics: B. Problem Statement
M q̈ + h = W λλ + W τ τ The objective of our real-time planning system is to connect
a C 2 -smooth, stable CoM trajectory for the next steps to the
LI˙ + RI + kM ω rot = U
current trajectory. To circumvent the instabilities of (9), we
B d˙ + Kd = f avoid solving an initial value problem, and solve a boundary
664 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 14, NO. 6, DECEMBER 2009

Fig. 9. Proposed walking control scheme.

Fig. 8. Example solution using spline collocation with 30 equally spaced A common strategy for stabilizing the upper body inclination
control points. ly := T x /Fz is the center of pressure. is to control the contact torques at the feet [9], [24], [25]. Usu-
value problem (BVP) for the time interval t ∈ [tB , tE ] instead ally, the foot torques are measured by a six-axis force/torque
sensor and controlled via position control of the ankle joints.
mb zb ÿb − mb yb (z̈b + g) = − T̃x (10) Another common strategy is to accelerate the CoM [2] and [26],
creating a reaction force that stabilizes the robot. In order to
(yb (tB ), ẏb (tB ), yb (tE ))T =(yb,B , ẏb,B , yb,E )T . (11)
reduce landing impacts, many robots also incorporate an active
Obviously, this is an ill-posed problem, since three boundary control of vertical contact forces. This can be done by measur-
conditions are prescribed for a second-order ODE. To obtain a ing the contact force and changing position set points [9], [24].
well-posed problem, we modify the reference torques by adding Large impacts from early foot contact have also been avoided by
γ∆Tx , where ∆Tx is a shape function, and γ the additional using low gain joint position control in combination with gain
degree of freedom required. scheduling [26] or inverse dynamics feed forward control [2].
These control components are often combined in an ad hoc
C. Numerical Solution manner. However, since they are not completely orthogonal, un-
Generally, (9) has no closed form solution, so we have imple- wanted interference is possible. Such unwanted interactions can
mented a collocation method with cubic splines as basis func- be avoided by using an integrated control approach. Löffler pro-
tions to numerically solve the BVP in real-time. An important posed an approach based on feedback linearization for the biped
property of cubic splines is their linearity in the control points JOHNNIE. Unfortunately, the performance of this method was
p, i.e., si (t) = ∇p si (t)p. Calculating the gradients ∇p si (t) in- limited by the available sensor bandwidth, computational power,
volves solving a tridiagonal system of linear equations, which model accuracy, and the difficulty performing joint torque con-
can be done very efficiently by LR-decomposition requiring only trol without joint torque sensors [9].
O(dim(p)) operations.
Putting all elements together, we have a linear set of equations B. Proposed Method
in the unknown spline parameters p We propose a generalization of the impedance control method
mb [zb,i ∇η̈i − ∇ηi (z̈b,i + g)]p = −T̃x,i − γ∆Tx,i (12) previously used for JOHNNIE [9]. The basic idea is to use hybrid
force/position control in task space with an inner joint position
pT (∇p η0T , ∇p η˙0 T , ∇p ηnT −2 ) = (yB , ẏB , yE ). (13) control loop. The contact force trajectories are modified in an
Here, η denotes the approximation of the solution of (10). outer control loop to provide inertial stabilization (Fig. 9).
Fig. 8 shows a solution obtained by the proposed method for The method uses a kinematic model of the robot describing
a periodic gait at 2.5 km/h using 46 cm steps and leg masses the dependency of the workspace trajectories x on the general-
ml of 8.2% of the robot mass. We used 30 spline parameters ized coordinates q. The workspace trajectories include relative
for three walking steps. Evidently the trajectories are smooth positions of the feet with respect to the CoM, and the orientation
and the contact torques stay within the region prescribed by the of the feet with respect to the upper body.
sequence of support polygons. Furthermore, we introduce a simplified contact model for
calculating contact forces λ as a function of q.
VIII. STABILIZING CONTROL 1) Contact Force Modification: If the reference contact force
trajectory λid from the on-line walking pattern generator is not
Due to modeling uncertainties, disturbances and an inherent modified, even small perturbations will destabilize the robot.
instability of the walking system, on-line stabilizing control of Following the idea described in Section VIII-A, desired contact
the robot is required. forces are modified in order to stabilize global dynamics. We
have chosen a potential difference (PD)-type control with sat-
A. Related Papers uration to obtain the modified contact torque reference λd as a
A large number of methods for stabilizing control of biped function of the torso’s orientation error ∆ϕtorso
robots have been proposed. A number of similar ideas have λd = sat(λ
λid + K P ∆ϕtorso + K D ∆ϕ̇torso ) . (14)
been successfully implemented on full-size humanoid robots.
Without attempting to review all methods, we briefly describe The orientation ϕtorso is expressed by the rotation about the
the ones we consider to be the most relevant to our approach. gravity vector ϕz , and the relative angles ϕx , ϕy between the
LOHMEIER et al.: SYSTEM DESIGN AND CONTROL OF ANTHROPOMORPHIC WALKING ROBOT LOLA 665

Fig. 11. Simulated upper body orientation for LOLA walking in a straight line
(ϕ x in sagittal, ϕ y in lateral plane).

Fig. 10. Simulation of LOLA and walking experiment with JOHNNIE, both
using the proposed control system. where (·)# denotes the Moore–Penrose pseudoinverse. ∇q λ is
calculated from (14) and will approach zero for the swing leg,
therefore, (17) is modified to include a position control term
torso’s vertical axis and the gravity vector in the sagittal and
lateral planes, respectively. K P and K D are gain matrices and
#
∆ẋλ = αλ { S λ ∇q λ(S̄ x ∇q x)# S λ K λ,ff λ λ̇d
sat is a saturation function that returns only physically feasi-

ble contact forces. The control gains are scheduled to follow +K λ (λλd − λ) } + αx S̄ x K λx (xd − x) . (18)
changing load factors during different walking phases and to
distribute the contact forces between left and right leg during Here, αλ and αx are gain matrices and αλ + αx = E is the
double support. unity matrix. The gain K λ,ff modifies the feed-forward term
2) Contact Model: The foot ground contact is modeled as a λ
λ̇d and is tuned during walking experiments. The force control
set of decoupled point contacts with stiffness C i and negligible gain αλ is zero for the swing leg, avoiding instability due to
damping. Using the known contact state, i.e., which contacts are a vanishing contact stiffness and allowing the position control
opened or closed, the total force F j and torque T j acting on term to move the swing leg back toward the reference trajectory.
foot j are given by To obtain a position control reference on joint level, the task
     space reference modified according to (18) must be mapped
Fj j C i dc,i into joint space. For this we use a resolved motion rate control
= (15)
Tj i∈I c , j j ∆r c,i × j C i dc,i scheme [28] with nullspace optimization [29]

where Ic,j denotes the subset of Ic on foot j, dc,i is the defor- Kx  

ẋd = ẋid + S̄ Tx ∆ẋλ + xid + S̄ Tx ∆xλ − x
mation of the ith contact element and j ∆r c,i is the vector from ∆t

the force sensor frame to the contact element.
3) Control Law: The subset of contact forces λc to be ac- ∆xλ = ∆ẋλ dt
tively controlled is set in a selection matrix S λ . We also select
a subspace xc = S x x to be position controlled using a second q̇ = (∇q x)# ẋ − αN N ∇q H. (19)
matrix S x . The use of selection matrices is similar to the use of Here, N is the nullspace projection matrix and αN a gain for
a selection vector proposed by Craig and Raibert [27] to choose minimizing the objective function H in x’s nullspace. Currently,
position and force controlled dimensions in task space. The ma- we use two components for H: joint limit avoidance and con-
jor difference lies in the fact that we have separate selection vergence to a “comfortable” pose. Joint limit avoidance is only
matrices for contact forces and task space dimensions, i.e., the activated in a boundary close to the joint limits and deactivated
selection of force- and position-controlled dimensions does not otherwise. For convergence toward a “comfortable” pose, we
have to be carried out in the same coordinate system and the use a quadratic cost function that penalizes divergence from a
robot may be redundant. reference pose according to each joint’s range of motion. The
In the nonposition-controlled dimensions, we allow a modifi- control method is described in more detail in [30].
cation of the reference trajectories ∆xλ for force control. With Fig. 10 shows snapshots of walking experiments with
K λ as a control gain, we choose the linear error dynamics JOHNNIE and a simulation of LOLA using the proposed plan-
  ning and control algorithms. An example of an IMU signal for
S λ ∆λ λ̇ + K λ ∆λ λ = 0. (16)
LOLA walking in our dynamics simulation is shown in Fig. 11.
With S̄ x denoting the complement of S x , we have
∆ẋλ = S̄ x ∇q x∆q̇. (17) IX. CONCLUSION

Using the least squares solution for ∆q̇ and λ λ̇ = ∇q λq̇, we Humanoid robots are a new and promising application area
obtain the following reference trajectory modification: for robotics. In order to be useful and commercially success-
 #   ful, humanoids must have reliable biped locomotion capabili-
∆ẋλ = (S λ ∇q λ S̄ x ∇q x )# S λ λ λ̇d + K λ (λ
λd − λ) ties. In this paper, we focused on aspects of the mechatronic
666 IEEE/ASME TRANSACTIONS ON MECHATRONICS, VOL. 14, NO. 6, DECEMBER 2009

hardware design and real-time control system of our new biped [20] R. Rockafellar, “Augmented lagrangians and applications of the proximal
robot LOLA. Its distributed electronics architecture, revised sen- point algorithm in convex programming,” Math. Oper. Res., vol. 1, no. 2,
pp. 97–116, 1976.
sor system, brushless motors, and extremely lightweight design [21] T. Buschmann, S. Lohmeier, M. Bachmayer, H. Ulbrich, and F. Pfeiffer,
significantly improve the physical capabilities. New trajectory “A collocation method for real-time walking pattern generation,” in Proc.
planning and control algorithms to improve walking capabilities IEEE-RAS Int. Conf. Humanoid Robot. (Humanoids), 2007, pp. 1–6.
[22] T. Takenaka, “Gait generation system of legged mobile robot,” Eur. Patent
were developed and extensively tested both in dynamics simu- 1 671 754-A2, Jun. 21, 2006.
lations and in walking experiments with JOHNNIE. LOLA has [23] J. Park and K. Kim, “Biped robot walking using gravity-compensated
been fully assembled, and we are currently performing initial inverted pendulum mode and computed torque control,” in Proc. IEEE
Int. Conf. Robot. Autom. (ICRA), 1998, pp. 3528–3533.
experiments. [24] S. Kajita, T. Nagasaki, K. Kaneko, K. Yokoi, and K. Tanie, “A running
controller of humanoid biped HRP-2LR,” in Proc. IEEE Int. Conf. Robot.
Autom. (ICRA), 2005, pp. 616–622.
ACKNOWLEDGMENT [25] T. Takenaka, “Controller of legged mobile robot,” Eur. Patent 1 475 198-
A1, Nov. 10, 2004.
The authors would like to extend sincere thanks to [26] K. Nishiwaki, “Design of walking system and online control of a humanoid
M. Bachmayer, V. Favot, G. Mayr, and M. Schwienbacher for robot,” Ph.D. dissertation, University of Tokyo, Tokyo, Japan, 2001.
their contributions to the LOLA project. The authors would also [27] J. Craig and M. Raibert, “A systematic method of hybrid position/force
control of a manipulator,” in Proc. Comput. Softw. Appl. Conf., IEEE
like to thank the reviewers for their helpful comments. Comput. Soc., 1979, pp. 446–451.
[28] D. E. Whitney, “Resolved motion rate control of manipulators and human
prostheses,” IEEE Trans. Man Mach. Syst., vol. MMS-10, no. 2, pp. 47–
REFERENCES 53, Jun. 1969.
[29] A. Liégeois, “Automatic supervisory control of the configuration and be-
[1] IFR Statistical Department. (2008). World robotics [Online]. Available:
havior of multibody mechanisms,” IEEE Int. Trans. Syst., Man, Cybern.,
http://www.worldrobotics.org/
vol. SMC-7, no. 12, pp. 63–71, Dec. 1977.
[2] D. Tajima, R. Honda, and K. Suga, “Fast running experiments involving a
[30] T. Buschmann, S. Lohmeier, and H. Ulbrich, “Biped walking control
humanoid robot,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), 2009,
based on hybrid position/force control,” in Proc. IEEE/RSJ Int. Conf.
pp. 1571–1576.
Intell. Robots Syst. (IROS), 2009.
[3] M. Hirose and K. Ogawa, “Honda humanoid robots development,” Philos.
Trans. R. Soc. A, vol. 365, pp. 11–19, 2007.
[4] K. Nishiwaki, J. Kuffner, S. Kagami, M. Inaba, and H. Inoue, “The ex-
perimental humanoid robot H7: A research platform for autonomous be-
haviour,” Philos. Trans. R. Soc. A, vol. 365, no. 1850, pp. 79–107, 2007.
[5] K. Kaneko, F. Kanehiro, S. Kajita, H. Hirukawa, T. Kawasaki, M. Hirata,
K. Akachi, and T. Isozumi, “Humanoid robot HRP-2,” in Proc. IEEE Int. Sebastian Lohmeier received the Diploma degree
Conf. Robot. Autom. (ICRA), 2004, pp. 1083–1090. in mechanical engineering from the Technical Uni-
[6] K. Kaneko, K. Harada, F. Kanehiro, G. Miyamori, and K. Akachi, “Hu- versity Munich, Garching, Germany, in 2003. He is
manoid robot HRP-3,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. currently working toward the Ph.D. degree at the In-
(IROS), 2008, pp. 2471–2478. stitute of Applied Mechanics, Technical University
[7] Y. Ogura, H. Aikawa, K. Shimomura, H. Kondo, A. Morishima, H. Lim, Munich.
and A. Takanishi, “Development of a humanoid robot WABIAN-2,” in He is a Research Assistant at the Institute of
Proc. IEEE Int. Conf. Robot. Autom. (ICRA), 2006, pp. 76–81. Applied Mechanics, Technical University Munich,
[8] J.-Y. Kim, I.-W. Park, J. Lee, M.-S. Kim, B.-K. Cho, and J.-H. Oh, “System where he has been engaged in the design and realiza-
design and dynamic walking of humanoid robot KHR-2,” in Proc. IEEE tion of the biped robot LOLA’s mechatronic system.
Int. Conf. Robot. Autom. (ICRA), 2005, pp. 1443–1448. His research interests include humanoid robotics,
[9] K. Löffler, M. Gienger, F. Pfeiffer, and H. Ulbrich, “Sensors and control mechatronic systems, actuators, and sensors.
concept of a biped robot,” IEEE Trans. Ind. Electron., vol. 51, no. 5,
pp. 972–80, Oct. 2004.
[10] S. Kajita, T. Nagasaki, K. Kaneko, K. Yokoi, and K. Tanie, “A hop towards Thomas Buschmann received the Diploma degree
running humanoid biped,” in Proc. IEEE Int. Conf. Robot. Autom. (ICRA), in mechanical engineering from the Technical Uni-
2004, pp. 629–635. versity Munich, Garching, Germany, in 2003. He is
[11] D. Kerrigan, U. Croce, M. Marciello, and P. Riley, “A refined view of currently working toward the Ph.D. degree at the In-
the determinants of gait: Significance of heel rise,” Arch. Phys. Med. stitute of Applied Mechanics, Technical University
Rehabil., vol. 81, pp. 1077–1080, 2000. Munich, where he has been engaged in the simula-
[12] K. Nishiwaki, S. Kagami, Y. Kuniyoshi, M. Inaba, and H. Inoue, “Toe tion and control of the biped robots JOHNNIE and
joints that enhance bipedal and fullbody motion of humanoid robots,” in LOLA.
Proc. IEEE Int. Conf. Robot. Autom. (ICRA), 2002, pp. 3105–3110. His research interests include walking robots,
[13] Design Methodology for Mechatronic Systems, VDI Std. 2206, Jun. 2004. multibody dynamics, and control.
[14] S. Kajita and K. Tani, “Experimental study of biped dynamic walking
in the linear inverted pendulum mode,” in Proc. IEEE Int. Conf. Robot.
Autom. (ICRA), 1995, pp. 2885–2891.
[15] S. Lohmeier, T. Buschmann, M. Schwienbacher, H. Ulbrich, and F. Pfeif- Heinz Ulbrich received the Ph.D. degree and the
fer, “Leg design for a humanoid walking robot,” in Proc. IEEE-RAS Int. Habilitation from the Technical University Munich,
Conf. Humanoid Robot. (Humanoids), 2006, pp. 536–541. Garching, Germany, in 1979 and 1986, respectively.
[16] S. Lohmeier, T. Buschmann, H. Ulbrich, and F. Pfeiffer, “Modular joint From 1979 to 1988, he was Chief Engineer at the
design for performance enhanced humanoid robot LOLA,” in Proc. IEEE Technical University Munich. After one year as Se-
Int. Conf. Robot. Autom. (ICRA), 2006, pp. 88–93. nior Research Fellow at the NASA Lewis Research
[17] J. Perry, Gait Analysis—Normal and Pathological Function, 3rd ed. Center, Cleveland, OH, he joined the Technical Uni-
Thorofare, NJ: Slack, 1992. versity Braunschweig as a Professor. He was a Pro-
[18] J. Bevans, “Biomechanics: A review of foot function in gait,” Foot, vol. 2, fessor at Essen University, Essen, Germany, during
no. 2, pp. 79–82, 1992. 1994. Since 2001, he has been the Head of the In-
[19] T. Buschmann, S. Lohmeier, H. Ulbrich, and F. Pfeiffer, “Dynamics simu- stitute of Applied Mechanics, Technical University
lation for a biped robot: Modeling and experimental verification,” in Proc. Munich. His current research interests include modeling, simulation, control,
IEEE Int. Conf. Robot. Autom. (ICRA), 2006, pp. 2673–2678. and realization of mechatronic systems.