Robust Motion Control of Oscillatory-Base Manipullators

Lecture Notes in Control and Information Sciences 463
Masayoshi Toda
Robust Motion
Control of
Oscillatory-Base
Manipulators
H∞-Control and Sliding-Mode-
Control-Based Approaches
Lecture Notes in Control and Information
Sciences
Volume 463
Series editors
Frank Allgöwer, Stuttgart, Germany
Manfred Morari, Zürich, Switzerland
Series Advisory Boards

P. Fleming, University of Sheffield, UK
P. Kokotovic, University of California, Santa Barbara, CA, USA
A.B. Kurzhanski, Moscow State University, Russia
H. Kwakernaak, University of Twente, Enschede, The Netherlands
A. Rantzer, Lund Institute of Technology, Sweden
J.N. Tsitsiklis, MIT, Cambridge, MA, USA
About this Series
This series aims to report new developments in the fields of control and information
sciences—quickly, informally and at a high level. The type of material considered
for publication includes:
1. Preliminary drafts of monographs and advanced textbooks
2. Lectures on a new field, or presenting a new angle on a classical field
3. Research reports
4. Reports of meetings, provided they are
(a) of exceptional interest and
(b) devoted to a specific topic. The timeliness of subject material is very
important.
More information about this series at http://www.springer.com/series/642

Masayoshi Toda
Robust Motion Control

of Oscillatory-Base
Manipulators
H1-Control and Sliding-Mode-Control-Based
Approaches
123
Masayoshi Toda
Department of Ocean Sciences
Tokyo University of Marine Science
and Technology
Tokyo
Japan
Additional material to this book can be downloaded from http://extras.springer.com.
ISSN 0170-8643 ISSN 1610-7411 (electronic)

Lecture Notes in Control and Information Sciences
ISBN 978-3-319-21779-6 ISBN 978-3-319-21780-2 (eBook)
DOI 10.1007/978-3-319-21780-2
Library of Congress Control Number: 2015944735
Mathematics Subject Classification: 70Q05
Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2016
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.
MATLAB® is a registered trademark of The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA 01760-
2098, USA, http://www.mathworks.com
Printed on acid-free paper
Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
Preface
The subject of the monograph is robust control of “An oscillatory-base manipulator

(OBM)” which can be regarded as a model system for a variety of mechanical
systems installed on the oscillatory base. Further, this class of systems can be
divided into subclasses, one of which mainly includes offshore mechanical systems,
e.g., offshore cranes, installed on the oscillatory base being affected by
wave-induced disturbances, and a typical example of the other subclass is a space
robot system mounted on the flexible base which tends to oscillate due to the
intrinsic flexible structure. This monograph will focus on the former subclass, i.e.,
offshore mechanical systems.
The monograph introduces such control problems, and presents some control
methodologies to solve them which the author has developed and demonstrates,
with respect to control system design and analysis, particularly on “robustness” and
control performances by simulations and hardware experiments. The common
feature of such control problems can be stated as “how to achieve successful
tracking control in the presence of disturbances due to the base oscillation and
further model uncertainties and variations.” Therefore, the model of an
oscillatory-base manipulator is a very important control objective because it serves
as a tool to solve various control application problems in the marine industries, but
also can contribute to the control science community as a benchmark system of
disturbance rejection and robust tracking control.
The control methodologies presented in the monograph are based on H1 control
and sliding-mode control, both of which are now well-known powerful control
schemes to solve robust control problems. The author believes that the control
problems considered are interesting and useful examples with respect to applica-
tions of those well-known control schemes. Therefore, the author strongly hope that
not only control engineers involved in marine systems, but also engineers in the
other areas and students in the course of control science get interested in such
problems, and control theory scientists will employ the problems to evaluate their
own novel control schemes.
v
vi Preface
Please note that the monograph has been designed to be as self-contained as

possible. Additionally, the reader can download some demonstration programs for
MATLAB® from http://www2.kaiyodai.ac.jp/*toda/obm/.
The author would like to express his sincere gratitude to Springer and the staff
who have given the author such an opportunity to publish the monograph and
helped a lot. And, the author wishes to express his thanks to his students who were
involved in this project and made the achievements together with the author, par-
ticularly, Mr. Masahiro Sato who supported him by preparing the estimation
algorithm-based programs in Chap. 7. Finally, the author wishes to express his deep
gratitude to his wife and children for their encouragement, support, and their
patience during the preparation of the monograph.
Tokyo, Japan Masayoshi Toda

June 2015
Contents
1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 What Is an Oscillatory-Base Manipulator? . . . . . . . . . . . . . . . . . 1
1.2 Previous Works and Contents of the Monograph . . . . . . . . . . . . 4
2 Problem Definition, Dynamical Model Formulation . . . . . . . . . . . . 7

2.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Local-Coordinate and Global-Coordinate Problems . . . . . . . . . . . 9
2.4 Dynamical Model Formulation. . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4.1 Dynamical Model of an OBM . . . . . . . . . . . . . . . . . . . . 10
2.4.2 Base Oscillation Model. . . . . . . . . . . . . . . . . . . . . . . . . 12
3 Experimental Apparatus and Analysis on Parameter Variation

Due to Payload . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Experimental OBM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 Analysis on Parameter Variation Due to Payloads . . . . . . . . . . . 17
4 Motion Control Using an H1 -Control-Based Approach . . . . . .... 21

4.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 21
4.2 H1 Control and μ-Analysis and Synthesis . . . . . . . . . . . . .... 22
4.2.1 Small-Gain Theorem and Linear
Fractional Transformations . . . . . . . . . . . . . . . . . . .... 23
4.2.2 H1 Control Standard Problems . . . . . . . . . . . . . . .... 24
4.2.3 Structured Uncertainties and μ-Analysis
and Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 26
4.2.4 TDOF Control System Structure . . . . . . . . . . . . . . .... 27
4.3 Extended Matrix Polytopes for Model
Uncertainty Representation . . . . . . . . . . . . . . . . . . . . . . . .... 28
4.3.1 Model Uncertainties and Extended Matrix Polytopes .... 28
4.3.2 LFT Representation of an Extended Matrix Polytope .... 30
vii
viii Contents
4.4 Control Design and Analysis . . . . . . . . . . . . . . . . . . . . . ..... 33

4.4.1 Nonlinear State Feedback and Virtual Linear Plant. ..... 34
4.4.2 Extended Matrix-Polytope-Based Model
Uncertainty Representation for the Inertia Matrix . . ..... 36
4.4.3 Sensitivity Function Shaping Strategy
Using Linear State-Feedback Control . . . . . . . . . ..... 41
4.4.4 Generalized Plant for H1 Control Design
and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . ..... 44
4.4.5 H1 Controller Synthesis and Robustness Analysis. ..... 51
4.5 Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ..... 66
5 Simulations and Experiments for the H1 -Control-Based

Approach. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2 Simulations and Experiments (Nominal-Case Performance) . . . . . 68
5.2.1 Base Oscillation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.2.2 Global Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.3 PID Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.4 Attitude Control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.5 Position Control. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.2.6 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.2.7 Influence of Sensor Error . . . . . . . . . . . . . . . . . . . . . . . 80
5.2.8 Experimental Results and Comparison
with PID Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.3 Simulations and Experiments (Robust Performance) . . . . . . . . . . 88
5.3.1 Robust Control Simulations and Experiments . . . . . . . . . 88
5.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.4 Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6 Motion Control Using a Sliding-Mode-Control-Based Approach . .. 97

6.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 97
6.2 Sliding-Mode Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 98
6.3 Sliding-Mode Control via Rotating Sliding Surface
with Variable-Gain Integral Control . . . . . . . . . . . . . . . . . . . . . 99
6.3.1 Control System Design of SMC-RSSI . . . . . . . . . . . . . . 99
6.3.2 Control System Design Example . . . . . . . . . . . . . . . . . . 102
6.4 Stability Analysis of RSSI. . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.5 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.5.1 Simulations in the Nominal Case . . . . . . . . . . . . . . . . . . 109
6.5.2 Simulations in the Robust Control Case . . . . . . . . . . . . . 118
6.6 Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Contents ix
7 Base Oscillation Estimation via Multiple H1 Filters . . . . . . . . . . . 123

7.1 Introduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.2 Estimation Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
7.2.1 FFT-Based Linear State-Equation Model. . . . . . . . . . . . . 125
7.2.2 Kalman Filtering Algorithm . . . . . . . . . . . . . . . . . . . . . 126
7.2.3 H1 Filtering Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 127
7.2.4 Method of Selectively Combining Multiple H1
and Kalman Filters. . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
7.2.5 Integral with Drift Error Compensation. . . . . . . . . . . . . . 129
7.3 Ship Oscillation Motions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.3.1 Ocean Wave Spectrum Model . . . . . . . . . . . . . . . . . . . . 130
7.3.2 Ship Dynamical Model . . . . . . . . . . . . . . . . . . . . . . . . . 130
7.4 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.4.1 Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
7.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
7.5 Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
References. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
Notation
Abbreviations
DOF Degree of freedom
FFT Fast Fourier transform
LFT Linear fractional transformation
LQ Linear quadratic
MIMO Multi-input multi-output
OBM Oscillatory-base manipulator
PID Proportional integral derivative
PSD Power spectral density
RMSE Root-mean-square error
RSSI Rotating sliding surface with variable-gain integral control
SDC State-dependent coefficient
SISO Single-input single-output
SMC Sliding-mode control or controller
TDOF Two-degree-of-freedom
VSS Variable structure system
Symbols
jj jj Euclidean norm
jj jjF Frobenius norm
jj jj1 H1 norm
ðÞT Transpose
R The field of real numbers
Rn The n-dimensional real vector space
Rnm The n by m real matrix space
C The field of complex numbers
Cnm The n by m complex matrix space
H1 The stable transfer function matrix space
xi
xii Notation
0 Positive definite
0 Negative definite
I The identity matrix
In The n by n identity matrix
diagðÞ Diagonal matrix
blockdiagðÞ Block diagonal matrix
σ ðÞ The largest singular value
F l ð; Þ Lower LFT
F u ð; Þ Upper LFT
Sð; Þ Redheffer star-product
S The sensitivity function matrix
T The complementary sensitivity function matrix
Ta The quasi-complementary sensitivity function matrix
μ The structured singular value
μc The μ with constant scalings
ðÞn Nominal parameter
PðÞ Orthogonal projection
vecðÞ Mapping from R22 to R4
vec1 ðÞ Inverse mapping of vec
1 The criterion 1 for the hyperplane
2 The criterion 2 for the hyperplane
e Root-mean-square error (RMSE)
Chapter 1
Introduction
Abstract This introductory chapter first defines the key terminology “oscillatory-
base manipulator,” which represents a model system associated with mechanical sys-
tems installed on the oscillatory base. This category consists of two groups depending
on the presence or the absence of external oscillatory disturbance. Typical examples
with external oscillatory disturbance can be found in marine mechanical systems,
such as offshore cranes, drill ships. The other group, i.e., with no external oscillatory
disturbance, mainly contains space robots mounted on the flexible base. This mono-
graph will focus on the former group of marine mechanical systems. We introduce
related works and our own research history where the H∞ control framework and the
sliding-mode control one have played important roles, and present the organization
of the monograph.
1.1 What Is an Oscillatory-Base Manipulator?
“An oscillatory-base manipulator (OBM)” which we will focus on throughout this

monograph is a model system of a variety of mechanical systems installed on the
oscillatory base. In particular, typical examples can be found in the marine engi-
neering and science fields, such as offshore cranes, drill ships, marine observatory
systems, on-board radar gimbal systems, and so on, which are subject to wave-
induced base oscillation. In this monograph, we will address motion control problems
for oscillatory-base manipulators by analyzing their dynamics, developing control
design methods, and demonstrating our proposed control systems.
The common feature of such control problems can be stated as “how to achieve
successful tracking control in the presence of disturbances due to the base oscillation
and further model uncertainties and variations.” Additionally, the desired trajectory
for the system to track is strongly affected by the base oscillation in many control
application scenarios. Therefore, the model of an oscillatory-base manipulator is a
very important control objective because it serves not only as a tool to solve vari-
ous control application problems in the marine industries but also as a benchmark
system of disturbance rejection and robust tracking control for the control science
© Springer International Publishing Switzerland 2016 1

M. Toda, Robust Motion Control of Oscillatory-Base Manipulators,
Lecture Notes in Control and Information Sciences 463,
DOI 10.1007/978-3-319-21780-2_1
2 1 Introduction
community. Due to the nature of the problems, additional requirements are neces-
sary for the control systems for OBMs compared to the conventional robot control
systems, e.g., [1, 8, 19, 53, 77, 89–91].
In general, control problems of OBMs can be categorized into two classes depend-
ing on the presence or absence of external oscillatory disturbance. One class of them
contains problems in the presence of external oscillatory disturbance which is the
main cause of base oscillation, to which offshore mechanical systems as described
above and mobile robots with manipulators running on a rough road belong [35, 36,
46, 64, 110]. The other class is of control problems in the absence of such external
disturbance, where, due to the intrinsic flexible structure of the system, motion of the
manipulator induces the base oscillation, for instance, a large space robot mounted
on a flexible base or free-flying one maintaining a space satellite is an example.
The latter class of problems have gained a lot of research interests in the space
engineering field. Manipulator systems involved in these works are referred to as
“macro/micro manipulators” [18, 74, 87, 99]. The central issues of such problems
are first to decouple the manipulator and base dynamics and second to decay the base
oscillation. To satisfy those requirements, several approaches have been proposed
such as task space feedback [36, 87], filtering command [65], path-planning [70, 79],
acceleration feedback [57], active damping [58], H∞ control [96, 97], Fuzzy control
[61], pseudo-inverse jacobian [62]-based approaches.
It should be noted that these two types of OBM control problems are essentially
different ones due to difference in the mechanisms of base oscillation arising. There-
fore, requirements for the respective control systems are also different. In the former
problems, the control system is expected to cope with the external oscillatory distur-
bance, in other words, to be able to excite the system according to the disturbance.
In theory, so called the “internal model principle” comes in, and the requirement for
the linear time-invariant control system is equivalent to having the corresponding
poles to the frequencies of disturbance. On the other hand, in the latter problems,
conversely the control system need not to excite the intrinsic oscillatory modes. If
the control system is linear time-invariant one, this requirement can be achieved by
letting the controller have the corresponding zeros to the oscillatory modes. In [97],
we have realized the requirement by utilizing H∞ control. Interestingly enough,
these two problems form a contrast to each other, in terms of system poles and zeros
for the controllers.
In this monograph, we focus on the former class of problems and furthermore
consider robust control against model uncertainties due to the payload variations,
which is often the case for oceanic industries such as the fishing industry and the ocean
mining industry where a variety of objects in size, shape, and mass are to be handled.
In recent years, there have been increasing requirements for offshore mechanical
systems such as high operability, accuracy, efficiency, and safety of operation, as the
role of the ocean has becoming more and more important in terms of energy, mineral,
and food resources, and further the global environmental problems.However, offshore
1.1 What Is an Oscillatory-Base Manipulator? 3
environments are considerably difficult for mechanical systems because of wave,

tide, and wind, which make the control problems of offshore mechanical systems
challenging.
Here, we introduce some of the previous works on control of offshore mechan-
ical systems related to the problem of OBMs. This class of control problems of
OBMs contains two types of problems according to the coordinate systems, that is,
base-fixed coordinates (local coordinates) and earth-fixed coordinates (global coor-
dinates). Most of the previous works considered global coordinates. As the most
popular control problem in such problems considering global coordinates, heave
(vertical) motion compensation systems have been extensively studied aiming at
applications in ocean platforms for deep water oil exploitation [2, 7], drill-ships
[15, 52], offshore cranes [21, 38, 45, 54, 71, 73], and remotely operated vehicles
[34]. The operations of the control systems in the ocean engineering literature can
be classified into two types, one of which is an operation to reduce oscillations of
ships or ocean platforms [2, 7], and the other one is to compensate their oscillatory
motions [15, 21, 38, 45, 54, 52, 67, 71, 73]. The control operation considered in the
monograph belongs to the latter type.
Then, we address the related works from the viewpoint of control theory. Being
based on the assumption that the dynamical model is almost linear, a PD-control-
based approach has been presented in [52], and LQ optimal control has been
employed in [2, 7, 38]. In [45], targeting on wire velocity control of an offshore
crane, a method of generating a reference signal for wave synchronization has been
proposed using a transfer function model. Skaare and Egeland [88] have presented a
parallel force/position control scheme with a similar transfer function model to that
in [45]. Do and Pan [15] have developed a nonlinear control system with a distur-
bance observer and a back-stepping technique. Neupert et al. [71] have exploited
the scheme of disturbance decoupling in [40] for a nonlinear model of a offshore
crane. In [67], a nonlinear internal-model-based approach has been developed for
autonomous vertical landing of an aircraft, whose dynamics is not affected by oscil-
latory disturbances. Except [52], in these papers, their proposed control systems have
been evaluated by performing numerical simulations as in [2, 7, 15, 38, 67] and/or
real model experiments [45, 71, 88].
The common objectives in such control problems of compensating oscillatory
motions are first tracking control according to the desired trajectory, which is strongly
influenced by the base oscillation, and second suppressing disturbance due to the base
oscillation. It is important to note that these control objectives can be interpreted into
the frequency-domain control objectives as long as the system can be regarded as
a linear time-invariant one, and the frequency range of the base oscillation can be
known in advance. In fact, which is the case for an offshore OBM, and hence we
considered application of H∞ control to such problems. H∞ is well known to be
useful and effective for such frequency-domain control objectives and moreover
robust control problems.
4 1 Introduction
1.2 Previous Works and Contents of the Monograph
First, we began with the simplest problem of an OBM in [94], where an H∞ -control-
based approach to motion control in local coordinates, i.e., base-fixed coordinates,
was examined for a one degree-of-freedom (DOF) manipulator with a one-DOF
base, thus a single-input single-output (SISO) system. The key idea in [94] as the
control methodology was collaboration of H∞ control, linear state-feedback con-
trol, and nonlinear state-feedback control. This method was extended to problems
of multiple-DOF manipulators in local coordinates, i.e., a multi-input multi-output
(MIMO) system [95] and further its robust control feature was enhanced by develop-
ing a machinery, the “extended matrix polytope” and by utilizing µ-synthesis based
on the machinery. Then, in [82], we applied the method to the global-coordinate
problems of OBMs by extending the control design method for local-coordinate
problems. Until the work [82], in order to evaluate the proposed method, we had
performed simulations and confirmed its effectiveness. After that, we developed an
hardware experimental apparatus which consists of a two-DOF manipulator and a
one-DOF base. In [83], we improved the method by employing a new weighting
function for H∞ control design, which not only reduces influence of sensor error
but also enhances the robustness of the control systems; furthermore the proposed
method was evaluated by simulations and hardware experiments. The results showed
that the proposed control system can successfully perform in tracking control while
suppressing disturbance due to the base oscillation, further even in the presence of
model uncertainties due to the payload variations.
Meanwhile, we attempted to apply sliding-mode control (SMC), which is also
known to be powerful to disturbance rejection and robust control problems [5, 20, 39,
41, 60, 73, 102, 103, 111]. In [42], we proposed a novel nonlinear sliding surface with
the variable-gain integral control. The proposed SMC provides superior performance
to the conventional SMC in terms of transient performance and less required control
inputs. The advantage of the SMC-based approach over the H∞ -control-based one is
that it does not require information of frequencies of the base oscillation. On the other
hand, its disadvantages are common shortcomings of SMC, that is, the gaps between
in theory and in practical applications mainly due to discontinuous switching control
inputs. In [42], only preliminary results have been given.
In those works, we assumed that accurate measurements of the base oscillation
are available for the control system. In [84], we considered estimation problem of the
base oscillation via a low-cost gyro rate sensor, and proposed a method of selectively
combining multiple H∞ filters. The important idea in [84] is a new criterion to select
appropriate filters, which is based on innovations. The simulation results showed that
the algorithm is useful but still needs to be improved.
By being based on our research history of robust control of OBMs and adding new
research results, we have completed this monograph. In this monograph, we have
reorganized and integrated the proposed methodologies and given a new perspective
to them. The chapters of the monograph are organized as follows.
1.2 Previous Works and Contents of the Monograph 5
• Chap. 2 presents the problem definition, including important assumptions, the

dynamical model formulation, and the base oscillation models, on which we will
be based throughout the monograph.
• Chap. 3 addresses the experimental apparatus which consists of a two-DOF manip-
ulator and a one-DOF base. With this apparatus, we are able to conduct robust
control tests by changing the attached payloads of 11 different weights. We also
analyze how these payload variations result in parametric variations in the dynam-
ical model of the manipulator.
• Chap. 4 presents the key content in our research history, that is, the H∞ -control-
based approach. First, we introduce our machinery the “extended matrix polytope”
to efficiently and less conservatively represent model uncertainties due to the pay-
load variations. Then, we explain and demonstrate the control design method.
Finally, we analyze the designed controllers with respect to the fundamental prop-
erties of the controller, i.e., poles and zeros, and frequency responses, and further
robustness.
• Chap. 5 demonstrates evaluations on the designed controllers in Chap. 4 by simu-
lations and experiments, which include robust control tests and comparison with
the conventional PID controller.
• Chap. 6 presents the SMC-based approach and evaluates the method by simula-
tions, which also include robust control tests and comparison with the proposed
H∞ -control-based approach.
• Chap. 7 introduces the estimation algorithm for the base oscillation, which deploys
a novel idea of selectively combining multiple H∞ filters. By simulations, we
demonstrate its estimation performance.
Chapter 2
Problem Definition, Dynamical Model
Formulation
Abstract This chapter presents the problem definition and dynamical model formu-
lation of an oscillatory-base manipulator considering an illustrative example, which
we will work on throughout the monograph. We will consider three types of control
problems, specifically attitude control in local coordinates (base-fixed coordinates)
which is associated with on-board operations such as cargo handling, attitude con-
trol in global coordinates (earth-fixed coordinates), e.g., radar gimbal systems, and
position control in global coordinates, e.g., heave-motion-compensated cranes. Fur-
ther, as patterns of base oscillation, we consider three patterns, single-frequency
sinusoidal oscillation, double-frequency sinusoidal one, and ocean-wave imitated
oscillation based on the Bretschneider spectrum. Using combinations of those cases,
we will demonstrate control system design and analysis, control simulations and
experiments.
2.1 Introduction
First, this chapter defines motion control problems of an OBM which we will work
on throughout this monograph, by choosing an illustrative model and setting some
important assumptions. Then, the problems are categorized into two classes, local-
coordinate problems and global-coordinate ones, depending on the coordinate frame
to be referred by the control system. We consider three types of control problems.
Subsequently, being based on the problem definition, the corresponding dynamical
model is derived. Further, we introduce three patterns of base oscillation which we
will use for various demonstrations together with the control problems.
2.2 Problem Definition
As an illustrative model of OBMs, we consider a two-DOF manipulator with a one-

DOF oscillatory base and a payload as depicted in Fig. 2.1, and motions of which
are restricted on the vertical plane. The reason that this simple model has been
DOI 10.1007/978-3-319-21780-2_2
8 2 Problem Definition, Dynamical Model Formulation
Fig. 2.1 Schematic diagram Y

of the illustrative model of
q2 a2
OBMs (a two-DOF payload
manipulator with a one-DOF
oscillatory base and a
payload)
Link 2 y q1 l1
qb Link 1
a1
gravity
Ly
o x
X
base O
Lx
employed is because, in practical cases of offshore mechanical systems, roll and

heave (vertical) oscillations among six-DOF ones are manifest and critical. It should
be noted that the roll motion of the model with respect to the origin is one-DOF
rotating one, however, with respect to another point, e.g., the root of the manipulator,
involves heave and sway motions as well as roll ones. Further, from the viewpoint
of control, forces parallel and torques orthogonal to the rotating axes of manipulator
joints do not affect the manipulator dynamics. Therefore, this illustrative model is
simple for exposition, but is realistic enough. Further, notice that a payload is attached
to the tip of the second link in Fig. 2.1, which is assumed to have uncertainty in its
physical parameters, e.g., mass and inertia moment, in order to analyze robust control
problems.
Here the following important assumptions are made in addressing motion control
problems of OBMs;
A1 the frequency range of the base motion is known in advance;
A2 the forces and torques exerted on the base by the manipulator are negligible to
the base motion;
A3 the actuator dynamics is negligible;
A4 the payload has uncertainty in its physical parameters, e.g., mass and inertia
moment;
A5 except the payload, the physical parameters of the manipulator are known;
A6 the joint angles and velocities of the manipulator can be measured; and
A7 the base oscillation angle can be measured.
A1 is natural and is not restrictive considering the ocean environment and, further, is
very important to make our proposed H∞ -control-based technique highly effective.
2.2 Problem Definition 9
In fact it has been reported that the frequencies of ocean surface waves are within the
range 1/30–1 Hz [9]. A2, which supports A1 together with the facts in oceanography
and also makes the control problem simple, shall be discussed later in detail and
be clarified not to be restrictive, but to be reasonable. A3 is only for simplicity. A4
is intended to analyze robust control problems, and is often the case for offshore
mechanical systems such as a crane handling mineral and fishery resources. A5 and
A6 are necessary for control systems we will demonstrate and standard conditions
in terms of motion control. These assumptions are taken into account in Chaps. 2–6,
and Chap. 7 which presents an estimation method of the base oscillation aims at
implementing A7.
Taking into account all the above conditions, now the motion control problem of
OBMs to be dealt with can be stated as “under the assumptions A1–A7, synthesize an
controller that achieves successful motion control of the manipulator in the presence
of disturbance due to the base oscillation and model uncertainties in the payload
physical parameters.”
2.3 Local-Coordinate and Global-Coordinate Problems
See that in Fig. 2.1 two different coordinate frames are set; O X Y is an inertia frame
whose Y –axis is parallel to the direction of gravity and ox y is a frame attached to the
base whose origin o is fixed at O. Motion control problems of OBMs can be typically
categorized into two classes depending on the coordinate frame to be referred by the
control system. One class is of the case where the control system performs being
based on a base-fixed coordinate frame, i.e., a local-coordinate frame (as ox y in
Fig. 2.1), which is called the class of local-coordinate problems. The other class
contains global-coordinate problems where an earth-fixed coordinate frame, i.e., an
inertia frame (as O X Y in Fig. 2.1), is referred. The local-coordinate problems are
associated with practical applications with base-fixed task spaces, such as a ship-
mounted crane performing load/unload operations of cargo on the ship. On the other
hand, an on-board radar gimbals and a ship-mounted crane performing on land fall
into the class of global-coordinate problems.
The common feature between both the problems in global and local coordinates is
the disturbances due to the base oscillation. On the contrary, the difference between
them is in that in global coordinates the desired trajectory of motion must be generated
according to the base motion, while it is not the case in local coordinates where the
reference signals do not contain information on the base oscillation and further there
is no need to measure the base motion as shown later.
Moreover, regardless of coordinate frames and in general, motion control prob-
lems are divided into position control ones and attitude control ones. In the case of
OBMs, “position control” implies that the position of the payload is to be controlled,
while “attitude control” does controlling the attitudes of the links. Hence, by com-
bining the coordinate frame classes with these motion control types, four patterns of
motion control problems of OBMs can be considered as depicted Figs. 2.2, 2.3, 2.4
Base
Fig. 2.2 Position control in a base-fixed coordinate system (local case)
Base
Fig. 2.3 Attitude control in a base-fixed coordinate system (local case)
Base
Fig. 2.4 Position control in a global-coordinate system
Base
Fig. 2.5 Attitude control in a global-coordinate system
and 2.5. As seen from the figures, notice that, in local-coordinate problems, position
control, and attitude control are equivalent as long as the redundancy and/or multi-
plicity of solution of the joint space mapped to the given task space can be ignored,
whereas it is not the case for global-coordinate problems.
2.4 Dynamical Model Formulation
2.4.1 Dynamical Model of an OBM
The dynamics of n-rigid link manipulators with revolutionary joints subject to

m-DOF base oscillation can be described by the standard formulation for manipu-
lator dynamics plus the disturbance due to the base oscillation as follows:
2.4 Dynamical Model Formulation 11
M(q)q̈ + C(q, q̇)q̇ + D q̇ + G(q, q b ) + H (q, q̇, q̇ b , q̈ b ) = τ , (2.1)
where q ∈ Rn and q b ∈ Rm denote the position vectors of the manipulator links

and the base, respectively, and τ is the input torque vector; M(q) ∈ Rn×n is the
inertia matrix of the manipulator, which is a symmetric positive definite matrix;
C(q, q̇)q̇ ∈ Rn represents the centripetal and Coriolis torque depending only on
the states of the manipulator; D ∈ Rn×n is the damping coefficient matrix of the
manipulator joints, which is a positive definite constant diagonal matrix; G(q, q b ) ∈
Rn is the gravitational torque; H (q, q̇, q̇ b , q̈ b ) ∈ Rn represents the inertia torque
and the centripetal and Coriolis torque due to the base oscillation. G(q, q b ) and
H (q, q̇, q̇ b , q̈ b ) form the disturbance prescribed, which are nonlinearly coupled with
both the states of the manipulator and the base.
Now, we derive the specific formulation of (2.1) for the illustrative model in
Fig. 2.1 by applying Lagrangian mechanics. The notations used here are as follows.
qb , q1 , q2 Position angles of the base and the links as defined in Fig. 2.1
J1 , J2 Inertia moment of each link with respect to the centroid
m1, m2 Mass of each link
D1 , D2 Damping coefficient of each joint
L x , L y , l 1 , a1 , a2 Geometric parameters as defined in Fig. 2.1
τ1 , τ2 Control torque applied to each joint
g Gravitational acceleration
t Time variable
Then, each term in the model (2.1) can be explicitly represented by (2.2)–(2.6)
and the base oscillation is described by (2.7); (·)T denotes the transpose.

M11 M12
M(q) =
M21 M22
M11 = m 1 a12 + m 2 (a22 + l12 + 2a2 l1 cos(q2 )) + J1 + J2
M12 = M21 = m 2 (a22 + a2 l1 cos(q2 )) + J2
M22 = m 2 a22 + J2 , (2.2)

−m 2 a2 l1 sin(q2 )q̇2 (2q̇1 + q̇2 )
C(q, q̇)q̇ = , (2.3)
m 2 a2 l1 sin(q2 )q̇12
D = diag[D1 , D2 ], (2.4)
G(q, q b ) = [G 1 , G 2 ]T
G 1 = −{m 1 a1 sin(qb + q1 ) + m 2 (l1 sin(qb + q1 )
+a2 sin(qb + q1 + q2 ))}g
G 2 = −m 2 a2 sin(qb + q1 + q2 )g, (2.5)
H (q, q̇, q̇ b , q̈ b ) = [H1 , H2 ]T

H1 = {m 1 (a1 2 − a1 L x sin(q1 ) + a1 L y cos(q1 ))
+ m 2 (a22 + l1 2 − l1 L x sin(q1 ) + l1 L y cos(q1 )
− a2 L x sin(q1 + q2 ) + a2 L y cos(q1 + q2 )
+ 2l1 a2 cos(q2 )) + J1 + J2 }q̈b
+ {m 1 (a1 L x cos(q1 ) + a1 L y sin(q1 ))
+ m 2 (l1 L x cosq1 + l1 L y sinq1
+ a2 L x cos(q1 + q2 ) + a2 L y sin(q1 + q2 ))}q̇b2
− 2m 2 a2 l1 sin(q2 )q̇b q̇2
H2 = {m 2 (a22 + a2 l1 cos(q2 ) − a2 L x sin(q1 + q2 )
+ a2 L y cos(q1 + q2 )) + J2 }q̈b
+ m 2 (a2 l1 sin(q2 ) + a2 L x cos(q1 + q2 )
+ a2 L y sin(q1 + q2 ))q̇b2
+ 2m 2 a2 l1 sin(q2 )q̇b q̇1 . (2.6)
nω
qb = i=1 Aωi sin(ωi t + φi ) (2.7)
As in (2.7), the base oscillation is modeled as a linear combination of multiple

sinusoidal motions. Equation (2.6) indicates that the manipulator is strongly influ-
enced by the base oscillation and that the larger amplitude and angular frequency of
the base will induce the larger H (q, q̇, q̇ b , q̈ b ). Hence, to achieve desirable motion
control, how to overcome H (q, q̇, q̇ b , q̈ b ) is the central issue.
2.4.2 Base Oscillation Model
In this monograph, we have chosen three types of base oscillations in (2.7) for demon-
strations of control design and control system evaluation. The first one is a single-
frequency one with ω1 = 2π (rad/s), Aω1 = 10 (◦) , and φ1 = 0 (rad), and the second
one is a double-frequency one with ω1 = π, ω2 = 2π (rad/s), Aω1 = Aω2 = 5 (◦) ,
and φ1 = φ2 = 0 (rad). Furthermore, for more realistic demonstrations, continu-
ously distributed frequencies imitating an ocean-wave spectrum are also considered
for the third oscillation model. As in [15], we have employed the two-parameter
Bretschneider spectrum [22], and by approximating the spectrum each component
of base motion in (2.7) is generated as in the following.
ωmax − ωmin
ωi = ωmin + (i − 1) (i = 1, 2, · · · , n ω ) (2.8)
nω − 1
1.25 ω04 2 −1.25(ω0 /ωi )4
Sωi = A e (2.9)
4 ωi5 s
2.4 Dynamical Model Formulation 13
0
10
Power
−1
10
−2
10 −1 0 1
10 10 10
Frequency (rad/s)
Fig. 2.6 Power spectrum of the Bretschneider model

ωi2 ωmax − ωmin
Aωi = 2Sωi (2.10)
9.8 nω − 1
where ω0 , ωmin , and ωmax are the modal, minimum, and maximum frequencies,
respectively, and As is the significant roll amplitude. Given these parameters, the
spectral density Sωi and amplitude Ai can be obtained. The phase φi is given as a
random number between 0 and 2π with the uniform distribution.
To generate a base rolling motion, those parameters are set as ω0 = 0.24π, ωmin =
0.2π, ωmax = 0.4π (rad/s), n ω = 10, As = π (rad). With those parameters, the
power spectrum is displayed in Fig. 2.6, where the circles represent the employed
frequencies. Moreover, since our experimental apparatus is a small-scale model,
taking into account the scale effect, (2.7) is modified as
nω
qb (t) = i=1 Aωi sin(5ωi t + φi ), (2.11)
i.e., the frequencies are magnified by 5 times so as to make the disturbance affect
enough. We refer to the oscillation as the Bretschneider oscillation hereafter.
Chapter 3
Experimental Apparatus and Analysis on
Parameter Variation Due to Payload
Abstract We developed an experimental apparatus in order to evaluate control sys-

tems for oscillatory-base manipulators (OBMs) by hardware experiments. We call
the apparatus the “experimental OBM,” which was designed to be in accordance with
the problem definition described in Chap. 2, and moreover to accommodate robust
control demonstrations by exchanging the attached payloads of different weights.
In this chapter, we first introduce the experimental OBM. Then, considering the
illustrative model given in Chap. 2, and the experimental OBM, we analyze how
the parameters in the dynamical model vary according to the payload variations, the
results of which will be utilized in robustness analyses and robust control simulations
for the control systems in the sequel.
3.1 Introduction
This chapter introduces an apparatus that we developed for experimental evaluation

on control systems for OBMs, which is called “experimental OBM.” The experimen-
tal OBM was designed to accommodate payload variation assuming robust control
problems. Utilizing the illustrative model in Chap. 2, and the experimental OBM,
we analyze the parameter variation in the dynamical model of manipulator due to
payload variation.
3.2 Experimental OBM
In order to perform experimental evaluations on control systems for OBMs, we have

designed and developed an apparatus (experimental OBM) such that the illustrative
model in Fig. 2.1 can be implemented. Figure 3.1 shows a photo of the experimental
OBM which has the same configuration as that of the illustrative model in Fig. 2.1,
namely a two-DOF manipulator with a one-DOF oscillatory base. By conducting
some system identification experiments, we have obtained the physical parameters
of the manipulator as shown in Table 3.1, which are deployed for control designs,
DOI 10.1007/978-3-319-21780-2_3
16 3 Experimental Apparatus and Analysis on Parameter Variation Due to Payload
Fig. 3.1 Experimental

oscillatory-base manipulator
Table 3.1 Physical parameters of the experimental OBM

Parameter Value Unit Parameter Value Unit
J1 7.373 × 10−4 kg m2 J2 1.228 × 10−4 ∗ k gm2
m1 2.180 × 10−1 kg m2 6.030 ×10−2 ∗ kg
∗
a1 5.739 × 10−2 m a2 1.947 × 10−2 m
l1 8.000 × 10−2 m l2 6.000 × 10 −2 m
D1 3.183 × 10−4 Nms D2 6.074 × 10−5 Nms
Lx 1.000 × 10−1 m Ly 6.000 × 10−2 m
g 9.800 m/s2
∗ These values are ones of the nominal case with the payload width of 5 mm
Table 3.2 Specifications of the motors installed on the experimental OBM

Motor b Motor 1 Motor 2
Power (W) 250 90 4.5
Maximum continuous 924 113 17
torque (mNm)
Maximum torque 1165 262.5 26.9
(mNm)
control system analyses, and simulation demonstrations of control system perfor-

mance throughout the monograph. The payload attached to the tip of Link 2 is
exchangeable with other payloads of various widths, i.e., various masses, which is
intended for analyses of robust control. In Table 3.1, the parameter values of Link 2
are ones of the nominal case with the payload width of 5 mm.
The joints are driven by the respective DC servo motors, Motor b for the base,
Motor 1 for Link 1, and Motor 2 for Link 2, whose specifications are shown in
Table 3.2. According to the specifications, the control torques are restricted as τ1
within ±0.2625 Nm and τ2 within ±0.0269 Nm. The angles of the links and base q1 ,
3.2 Experimental OBM 17
q2 , and qb are measured via the attached encoders with the resolution of 0.18 ◦ , and
the angular velocities are estimated using the backward difference of the angle data.
The sampling period for measurement and control is 0.01 s.
3.3 Analysis on Parameter Variation Due to Payloads
This section analyzes parameter variation in the dynamical model of manipulator

by utilizing the illustrative model shown in Fig. 2.1 and the experimental OBM.
Note that this analysis is not exclusive for OBMs but for general manipulators with
payload variations. As mentioned in the previous section, the payload attached to the
experimental OBM is exchangeable. We have five types of payloads that have the
same circular cross section and the different widths of 1, 2, · · · , 5 mm, respectively.
By choosing a single or multiple payloads, it is possible to implement various cases
of payload mass and inertia moment. Figure 3.2 shows a photo of samples of the
payloads and Table 3.3 presents their respective widths and masses.
We demonstrate how the parameters in the dynamical model in (2.2)–(2.5) vary
due to changes of the payloads. The case of the payload with 5 mm width is regarded
as the nominal case. To facilitate the analysis, we rewrite the dynamical model as
follows.
Fig. 3.2 Sample of payloads

for the experimental OBM
Table 3.3 Widths and masses of the payloads for the experimental OBM
Width (mm) 1 2 3 4 5
Mass (g) 2.06 4.11 6.21 8.28 10.49
18 3 Experimental Apparatus and Analysis on Parameter Variation Due to Payload
In the inertia matrix M(q),
M11 = M1 + 2Rcos(q2 ),
M12 = M21 = M2 + Rcos(q2 ),
M22 = M2 ,
M1 = m 1 a12 + m 2 (a22 + l12 ) + J1 + J2 ,
M2 = m 2 a22 + J2 ,
R = m 2 a2 l 1 . (3.1)
Hence, the centripetal and Coriolis term is rewritten as:

−R sin(q2 )q̇2 (2q̇1 + q̇2 )
C(q, q̇)q̇ = . (3.2)
R sin(q2 )q̇12
The variation of mass of the payload from that in the nominal case is equal to
the variation of mass of Link 2. Thus, let Δm 2 denote the variation of mass. With
the subscript (·)n which denotes the nominal parameter, Δm 2 = m 2 − m 2n . Then,
considering the shape and configuration of the payload, the variations in a2 (position
of centroid) and M2 in (3.1) can be calculated by the following:
Δm 2 (l2 − a2n )
Δa2 = , (3.3)
m 2n + Δm 2

r12 + r22
ΔM2 = Δm 2 + l2 ,
2
(3.4)
2
where r1 is the inner radius and r2 is the outer radius of the circular cross section of
the payload. Further using (3.3) and (3.4), the variations in the other parameters can
be obtained as follows:

r12 + r22
ΔM1 = Δm 2 + l1 + l2 ,
2 2
(3.5)
2
ΔR = Δm 2 l1l2 . (3.6)
From the above, therefore, the variations in the respective terms in (2.2)–(2.5) are
represented by
ΔM(q) = M(q) − Mn (q)

= Δm 2 M (q)

M11 M12
= Δm 2 ,
M21 M22
r12 + r22
M11 = + l12 + l22 + 2l1l2 cos(q2 ),
2
3.3 Analysis on Parameter Variation Due to Payloads 19
Table 3.4 Parameter variations due to change of the payload

Payload Variation m 2 a2 M1 M2 R
width rate∗ (%) (×10−2 kg) (×10−2 m) (×10−4 kg m2 ) (×10−4 kg m2 ) (×10−5 kg m2 )
(mm)
0 −17 4.981 1.093 1.881 1.065 4.36
1 −14 5.187 1.288 1.901 1.142 5.34
2 −11 5.392 1.467 1.922 1.219 6.33
3 −7 5.602 1.637 1.944 1.297 7.34
4 −4 5.809 1.793 1.964 1.374 8.33
5 0 6.030 1.947 1.987 1.457 9.39
6 3 6.236 2.081 2.008 1.533 10.38
7 7 6.441 2.206 2.029 1.610 11.37
8 10 6.651 2.325 2.050 1.688 12.37
9 14 6.858 2.436 2.071 1.765 13.36
10 17 7.079 2.548 2.093 1.848 14.43
∗ Variation rate of m 2
r12 + r22
M12 = M21 = + l22 + l1l2 cos(q2 ),
2
r12 + r22
M22 = + l22 , (3.7)
2
ΔC(q, q̇)q̇ = C(q, q̇)q̇ − Cn (q, q̇)q̇

−l1l2 sin(q2 )q̇2 (2q̇1 + q̇2 )
= Δm 2 , (3.8)
l1l2 sin(q2 )q̇12
ΔG(q, q b ) = G(q, q b ) − G n (q, q b )
= [ΔG 1 , ΔG 2 ]T ,
ΔG 1 = −Δm 2 {l1 sin(qb + q1 ) + l2 sin(qb + q1 + q2 )}g,
ΔG 2 = −Δm 2 l2 sin(qb + q1 + q2 )g, (3.9)
where the terms with the subscript (·)n denote the corresponding terms with the
nominal parameters.
It should be noticed that all the variations of the respective terms in the manipu-
lator dynamical model except the damping term is completely linear with respect to
the variation of mass of the payload Δm 2 . This remarkable feature stems from the
special property and configuration of the payload considered here, that is,
• the inertia moment is proportional to the payload mass; and
• the position of centroid of the payload on the link is invariant.
However, even in practical cases similar situations might be expected. Hence, this
feature can be and should be exploited in robust control design as demonstrated in the
subsequent chapter. Table 3.4 shows the corresponding parameters to the respective
payload widths, which are employed for demonstrations of control design, control
system analysis, simulations, and experiments.
Chapter 4
Motion Control Using an H∞ -Control-Based
Approach
Abstract This chapter presents the key contents of the monograph, that is, the H∞ -
control-based approach for the OBM robust control problems. This control method-
ology consists of several schemes, specifically nonlinear state-feedback control to
reduce the nonlinearity, parametric model uncertainty representation, H∞ control
with weighting functions, and additional linear state-feedback control to compen-
sate for the H∞ control scheme. In particular, in order to less conservatively and
effectively represent parametric model uncertainties due to the payload variations,
we have developed a machinery the “extended matrix polytope,” which is an exten-
sion of the conventional matrix polytope, in the aim of representation of a non-convex
parameter space. In this chapter, we begin with brief review of the basic notion of H∞
control, and introduce the extended matrix polytope. Then, we present the control
design method with some design examples considering four types of H∞ controllers,
and perform analyses on the designed control systems in terms of properties such
as system poles and zeros, frequency response, and robustness by utilizing two dif-
ferent approaches, one of which is based on μ-analysis with the extended matrix
polytope, and the other one is a Lyapunov-theory-based one with a state-dependent
coefficient (SDC) form. From the results of analyses, the designed control system
reveals favorable properties in terms of disturbance rejection, tracking control, and
robustness. Hence, their practical control performances can be also expected, which
will be shown in the next chapter.
4.1 Introduction
This chapter presents the H∞ -control-based approach for the OBM robust control
problems aforementioned, which is the heart of the monograph. The characteristic
features of the problem of OBMs are first disturbance due to the base oscillation;
second model uncertainties due to the payload variation; and third the intrinsic non-
linearity of the robotic manipulator. The first feature can be interpreted into the
frequency-dependent property. The second one requires a control design method
with which one can cope with robustness. When these two features are accounted
for, the H∞ control framework can be naturally a powerful candidate. The H∞
DOI 10.1007/978-3-319-21780-2_4
22 4 Motion Control Using an H∞ -Control-Based Approach
control that we have employed is fundamentally for linear time-invariant systems.

However, by adding some modifications, the nonlinearity can be reduced and be
absorbed into the model uncertainty.
Therefore, the point is how to accommodate and cope with the disturbance, model
uncertainty, and nonlinearity. In this context, we have developed the method to
express parametric uncertainties efficiently and effectively such as uncertainties due
to the payload variation. This method was developed by improving “matrix poly-
tope,” thus we call the method “extended matrix polytope” which is a less conserva-
tive approach than the conventional one and naturally contains the conventional one.
We will demonstrate the effectiveness of the extended matrix polytope in control
design and robustness analysis.
This chapter is organized as follows. In the next section, we briefly review the
fundamental notions and tools of H∞ control and μ-analysis and synthesis, which
will be minimum requisite for us to explain our idea in the sequel. Then, Sect. 4.3
introduce the extended matrix polytope. Further, Sect. 4.4 presents the control design
method and analyzes the obtained control systems. Taking into account of the page
length, evaluations on the designed control systems by simulations and experiments
will be presented in the next chapter.
4.2 H∞ Control and μ-Analysis and Synthesis
In this section, we briefly review the H∞ control framework. The history of this
attractive control design methodology started with the works [113, 114] by Zames
in the early 1980s, which pointed out that, as criteria to evaluate control systems,
the H∞ norm of the transfer function matrix of a closed-loop control system is
important. Before that, for instance, the well-developed linear quadratic (LQ) optimal
control framework had been based on the H2 norm which represents a sort of average
performance over the frequency domain. While, the H∞ norm represents a sort of
the maximal gain over the frequencies, and thus is strongly related to the worst-case
performance. Then, this idea was connected with so-called “the small-gain theorem”
and has developed to be one of the most effective and popular robust control design
framework [50, 104]. By motivated Zames’s works, a lot of research efforts have been
devoted and provided important and useful achievements [16, 24, 28, 37, 44, 51, 80].
In principle, the H∞ control method is available for linear, time-invariant systems.
However, this methodology has been extendedly applied to nonlinear, time-varying
systems by elaborate modifications, exactly as in this monograph, and, on the other
hand, also in the theoretical field, the notion of H∞ norm is extended to nonlinear
system in a natural manner as the induced L2 -gain-based control.
4.2 H∞ Control and μ-Analysis and Synthesis 23
Fig. 4.1 Feedback +

configuration for the G1
+
small-gain theorem
+
+
G2
4.2.1 Small-Gain Theorem and Linear Fractional

Transformations
Here, we address the considerably important theorem which made the H∞ control
framework a powerful robust control design method, that is, “the small-gain theo-
rem”. We consider a feedback configuration of two systems as depicted in Fig. 4.1.
Let G 1 (s) and G 2 (s) denote the transfer function matrices of the respective systems
in Fig. 4.1. Then, for the closed-loop system, the following theorem is known as the
small-gain theorem.
Theorem 4.1 (Small-Gain Theorem) [14, 30, 115] Suppose G 1 (s) ∈ H∞ and
G 2 (s) ∈ H∞ in the feedback configuration in Fig. 4.1. If and only if
||G 1 (s)G 2 (s)||∞ < 1 and ||G2 (s)G1 (s)||∞ < 1, (4.1)
the closed-loop system is internally stable.
In the robust control design approach, one of the two systems in Fig. 4.1 is desig-
nated as the uncertainty, and the other one plays a role of the interconnected system
consisting of the plant, controller, even fictitious weighting functions for design.
Then, in order to express such a configuration in design and analysis, the follow-
ing formula, called “linear fractional transformation (LFT)”, is useful and is com-
monly used.
Consider a block diagram in Fig. 4.2, which usually represents the plant including
the uncertainty denoted by Δ in modeling and/or control design. The block M con-
tains the nominal plant (i.e., when Δ = 0) and maybe additional weighting functions
reflecting the design specifications. Thus, M is in general referred to as a “general-
ized plant”. M’s input port for u and output port for y are usually used for a feedback
controller to be designed. Let Δ and M represent the system matrices, then according
to the respective dimensions M can be partitioned as

M11 M12
M= , (4.2)
M21 M22
Further, If (I − M11 Δ) is invertible, the upper LFT [17, 30, 115] can be defined as
Fu (M, Δ) := M22 + M21 Δ(I − M11 Δ)−1 M12 , (4.3)

Fig. 4.2 Upper LFT

configuration Δ
w z
u M y
Fig. 4.3 Lower LFT w z

configuration
u M y
which represents the transfer function matrix from u to y in Fig. 4.2.

On the other hand, one may consider a upside-down configuration as in Fig. 4.3 in
robustness analysis and/or problem formulation for H∞ control, where K represents
the feedback controller. In this configuration, M’s input port for w and output port for
z are used to evaluate the robustness of the closed-loop control system. Being based
on the same partitions of M in (4.2), if (I − M22 K ) is invertible, this configuration
is given by
Fl (M, K ) := M11 + M12 K (I − M22 K )−1 M21 , (4.4)
which represents the transfer function matrix from w to z in Fig. 4.3, and is called
the lower LFT.
4.2.2 H∞ Control Standard Problems
Using the system configuration in Fig. 4.3 and the lower LFT, the H∞ optimal control
problem can be formulated as
(H∞ optimal control problem) “Find the controller such that Fl (M, K ) is inter-
nally stable and ||Fl (M, K )||∞ is minimized.”
In practice, instead of the optimal control problem, the following H∞ suboptimal
control problem may be considered:
(H∞ suboptimal control problem) “ For given γ > 0, find a controller such that
Fl (M, K ) is internally stable and ||Fl (M, K )||∞ < γ .”
Moreover, we introduce a class of problems, called the “H∞ control standard

problem.” As the name suggests, this class of problems include a lot of general control
problems in practices, further it is well known that even other types of problems can
be reduced to this problem with some modifications. In fact, the problem for OBM
control falls into this category. For the standard problems, the solution techniques
have been well developed and the corresponding software products are available, for
instance, MATLAB® , Robust Control Toolbox.
Consider the state-space formulation to represent the generalized plant M in
Fig. 4.3 as follows:
ẋ = Ax + B1 w + B2 u, (4.5a)
z = C1 x + D11 w + D12 u, (4.5b)
y = C2 x + D21 w + D22 u, (4.5c)
where x is the state, u is the control input, y is the output and fed back to the
controller, w and z are the fictitious input and output for evaluation in the sense of
H∞ norm, with compatible dimensions respectively.
Then, we assume the following:
(A1) (A, B2 ) is stabilizable and (C2 , A) is detectable.
(A2) D12 is full column
rank and D21 is full row rank.
A − iωI B2
(A3) has full column rank for ∀ω ∈ R.
C1 D12

A − iωI B1
(A4) has full row rank for ∀ω ∈ R.
C2 D21
The solvability condition for the problem and the specific formulation of the
resultant controller are not presented in this monograph, which are not necessary
to understand the essential ideas and the story in this monograph. But interested
readers can refer to, e.g., [16, 28, 30, 115] for details. There exist two typical and
popular approaches to solve the H∞ control problems, one of which is the one to
solve two algebraic Riccati equations, so-called “γ -iteration” [16, 28], and the other
one is based on linear matrix inequalities (LMIs) [44, 80]. Both the algorithms can
be available by using MATLAB® , Robust Control Toolbox.
Remark 4.1 If Fl (M, K ) is interconnected with the uncertainty model Δ in the

feedback configuration, then the H∞ control problem becomes the robust control
problem in the presence of Δ. Further, if Δ is stable, that is, Δ ∈ H∞ and ||Δ||∞ <
1/γ , then according to Theorem 4.1 (small-gain theorem) the stabilizing controller
K such that ||Fl (M, K )||∞ ≤ γ ensures the robust stability of the entire control
system, which is summarized in the following theorem:
Theorem 4.2 (Robust stability for unstructured uncertainties) Let the feedback con-
trol system Fl (M, K ) be stable and be interconnected with the stable uncertainty sys-
tem Δ(s) ∈ H∞ bounded as ||Δ(s)||∞ < 1/γ (γ > 0) in the feedback configuration.
The entire closed-loop system is robustly stable, if and only if ||Fl (M, K )||∞ ≤ γ .
4.2.3 Structured Uncertainties and μ-Analysis

and Synthesis
If the uncertainty model matrix Δ is not given any specific structure, then it is called
a full matrix and an “unstructured uncertainty”. Otherwise, Δ is called a “structured
uncertainty”. In the case of structured uncertainties, the controller K mentioned in
Remark 4.1 can be conservative. “The structured singular value” theory [4, 17, 72, 92]
was developed to accommodate such cases in both robustness analysis and control
design.
Without loss of generality, the structure of square uncertainty Δ ∈ Cn×n is given
by the set as

= blockdiag[δ1 Ir 1 , . . . , δs Ir s , Δ1 , . . . , Δ f ] : δi ∈ C, Δ j ∈ Cm j ×m j , (4.6)
s f
where i=1 ri + j=1 m j = n. Then, the structured singular value with respect to
is first defined for a constant matrix M ∈ Cn×n as in the following.
1
μ (M) := , (4.7)
min{σ̄ (Δ) : Δ ∈ , det(I − MΔ) = 0}
where σ̄ (·) denotes the largest singular value. If there exists no Δ ∈ which makes
the matrix singular, then μ (M) := 0. Furthermore, this notion can be extended to
a linear time-invariant system matrix M(s) ∈ Cn×n as
μ (M(s)) := sup μ (M(iω)). (4.8)

ω∈R
The method to analyze robustness of the system using μ instead of ||M(s)||∞ =

supω∈R (σ̄ /M(iω)) (when M(s) ∈ H∞ ) is called “μ-analysis”. As described above,
when evaluating the robust stability of the feedback configuration of the structured Δ
and Fl (M, K ), the method using only their H∞ norms will provide a conservative
result. Hence, the μ-analysis is a less conservative approach in such a case, and
the counterpart of Theorem 4.2 for structured uncertainties is expressed as in the
following:
Theorem 4.3 (Robust stability for structured uncertainties) Let the feedback control
system Fl (M, K ) be stable and be interconnected with the stable uncertainty system
Δ(s) ∈ H∞ bounded as ||Δ(s)||∞ < 1/γ (γ > 0), which has a structure that
Δ(iω) ∈ f or ∀ω ∈ R, in the feedback configuration. The entire closed-loop system
is robustly stable, if and only if μ (Fl (M, K )) ≤ γ .
In general, μ cannot be directly computed. The alternative method to compute μ
is based on its upper and lower bounds. In particular, in practical applications, the
upper bound is utilized because it can provide the sufficient condition for robustness.
Therefore, here we introduce only the upper bound. Consider the following matrix
set where a matrix has a compatible dimension with Δ ∈ :

D = blockdiag[D1 , . . . , Ds , d1 Im1 , . . . , d f Im f ] : Di ∈ Cri ×ri , Di = Di∗ 0, d j > 0 ∈ R .
(4.9)
D ∈ D is a scaling matrix compatible with Δ ∈ . Then, the following inequality

always holds for a constant matrix M:
μ (M) ≤ inf σ̄ (DMD−1 ). (4.10)

D∈D
Next, we address a control design method based on μ called “μ-synthesis”, which

utilizes the μ upper bound in (4.10). In the μ-synthesis, considering Theorem 4.3,
the objective is to synthesize stabilizing a controller K and frequency-dependent
scalings D(iω)’s ∈ D such that

sup inf σ̄ DFl (M, K )D −1 (iω) < γ . (4.11)
ω∈R D∈D
This searching procedure is to iteratively repeat steps, and each step which consists
of two sub-steps such that K or D is alternately searched while keeping the other
one fixed. Therefore, this procedure is called the “D–K iteration.”
Remark 4.2 Note that the above μ-based approaches are available in theory for time-
invariant uncertainties. As described later, in the control problem of OBMs time-
varying uncertainties need to be dealt with. In the case of time-varying uncertainties,
the frequency-dependent scalings D(iω)’s have to be replaced by a constant scaling
over ∀ω ∈ R in both the μ-analysis and synthesis [3, 43, 68, 76, 79, 86, 93]. However,
in general, the D–K iteration using a constant D is difficult to find an acceptable
solution. Therefore, in this monograph, we will demonstrate the control design using
the conventional D–K iteration and the μ-analysis with a constant scaling. Many
research efforts have been devoted to the constant D–K iteration, e.g., [3, 79, 109].
4.2.4 TDOF Control System Structure
It is well known that H∞ control is generally a very powerful tool to achieve fre-
quency domain objectives, however, it has a shortcoming in that one cannot straight-
forwardly incorporate time domain objectives, such as overshoots, by using it alone.
Therefore, when necessary, a two-degree-of-freedom (TDOF) control system struc-
ture [10, 30, 75] is utilized together with the H∞ control. See Fig. 4.4 which shows a
typical TDOF control system structure. Pmdl represents the model function which has
a desirable time domain property, and the control system will try to let the closed-loop
system exactly match Pmdl [75].
Fig. 4.4 TDOF control

system structure; P actual Pn-1Pmdl
plant, Pn nominal plant, Pmdl
model plant, K controller, y +
output to be controlled, r + +
r Pmdl K P y
reference command _
4.3 Extended Matrix Polytopes for Model Uncertainty

Representation
4.3.1 Model Uncertainties and Extended Matrix Polytopes
In many applications of offshore mechanical systems, the payload variation is the

general case, which leads to the inertia parameter variation of the system as demon-
strated in Chap. 3, e.g., when dealing with mineral resources, fishery resources, etc.
In this section, we address the way of accommodating such model uncertainties for
H∞ control design and μ synthesis. The control design method via H∞ control pre-
sented in this chapter requires a special treatment to represent model uncertainties
in the inertia matrix. Hence, the objective here is to represent the uncertainties in an
LFT form so that the problem can be dealt with in the H∞ control framework. The
aim of the following argument is to represent model uncertainties in a compatible
manner with LFTs, as less conservatively as possible, depending on given informa-
tion about those uncertainties. First, the motivation is addressed and the proposed
methodology is outlined.
Suppose a matrix X ∈ R2×2 with two perturbational parameters Δx11 and Δx22
as described by the following:

x11 x12
X= = X n + ΔX,
x21 x22

x11n x12n
Xn = , ΔX = diag[Δx11 , Δx22 ], (4.12)
x21n x22n
where X n and ΔX denote the nominal matrix and the perturbational matrix which
represents uncertainties of X respectively. When one knows only the bounds of those
perturbations, |Δx11 | ≤ β1 and |Δx22 | ≤ β2 , the possible perturbational parameter
set is depicted by the solid lined Area 1 in Fig. 4.5, which shows the Δx11 –Δx22 space.
When designing robust controllers for those perturbations, one must consider all the
perturbations in Area 1. However, with more information about the dependence of
those, e.g., that perturbations only occur with a common sign, which implies that all
perturbed parameters simultaneously increase or decrease, the possible parameter set
can be restricted to a smaller area such as the dashed lined Area 2. When assuming
4.3 Extended Matrix Polytopes for Model Uncertainty Representation 29
Fig. 4.5 Various ranges on Δ x 22

the Δx11 –Δx22 plane
β2 ( β ’1, β 2)
( β 1, β ’2)
-β 1 β1 Δ x 11
(- β 1 , - β ’ 2 ) Area 1
Area 2
Area 3
(- β ’ 1 , - β 2 ) - β 2
the inertia matrix perturbations due to payload variation, in fact, which is often
the case, which has been already analyzed for the illustrative model in Chap. 3. In
accordance with the increase of the payload for the manipulator, all or some entries
in M(q) in (2.22) will increase and vice versa. Then it is obviously conservative to
consider all of Area 1 for such a case. Further with more detailed information, such
as dependency between the perturbational parameters, one can take more effective
design approaches as depicted by Area 3 in Fig. 4.5.
Which approach to be taken depends on the situations. In particular, when one
is dealing with payload variation cases, the approach associated with Area 3 can
possibly be taken. In the sequel, an appropriate method to represent the perturbation
space in an LFT is introduced. Then, in Sect. 4.4, by using the numerical example of
the illustrative model in Chap. 3, this notion will be demonstrated in detail.
One conventional tool to represent convex hulls such as Area 1, the matrix polytope
[31, 105], can be available. For example, a representation of Area 1 using a matrix
polytope is as in the following:

4
ΔX = λi X ai , (4.13)
i=1
where X a1 = diag[β1 ,β2 ], X a2 = diag[−β1 , β2 ], X a3 = −X a1 , X a4 = −X a2 , λi

4
∈ R, 0 ≤ λi ≤ 1 and i=1 λi ≤ 1. X ai ’s correspond to the respective vertices of
Area 1 and are therefore called vertex matrices.
Then, let us consider how to represent Areas 2 and 3. As seen from the figure,
these are not convex; hence matrix polytopes cannot be directly applied. However,
by introducing scaling parameters, this polytope approach can be easily extended so
as to accommodate cases such as Area 2 as follows:

3
ΔX = ν4 νi X bi , (4.14)
i=1
where X b1 = diag[β1 , 0], X b2 = X a1 , X b3 = diag[0, β2 ], νi ∈ R, 0 ≤ νi ≤ 1 for

3
i = 1, . . . , 3, i=1 νi = 1, and only |ν4 | ≤ 1, which representation is referred to as
an extended matrix polytope. Further, by using a completely identical manner, Area
3 is also represented as follows:

2
ΔX = κ3 κi X ci , (4.15)
i=1
where X c1 = diag[β1 , β2 ], X c2 = diag[β1 , β2 ], κi ∈ R, 0 ≤ κi ≤ 1 for i = 1, 2,

2
i=1 κi = 1, and only |κ3 | ≤ 1.
Thus, the problem is how to transform the above polytope and extended polytope
representations into LFTs, which is, in fact, possible. One systematic transformation
procedure for general cases is derived in the next section, and with which polytopes
and extended polytopes can be dealt with in a uniform manner.
4.3.2 LFT Representation of an Extended Matrix Polytope
In order to design and analyze control systems with model uncertainties in the H∞
control framework, it is required and powerful to represent the uncertainties in LFT
forms. To this end, here an LFT representation of an extended matrix polytope is
derived as in the following procedure: the uncertainty parameter set of an extended
matrix
3 polytope, which has an element-by-element dependent constraint such as
i=1 i = 1 in (4.14), is transformed into a parameter set with independent con-
ν
straints; for this parameter set an LFT is derived; then the range of this parameter set
is normalized by using another LFT; finally an LFT of the extended matrix polytope
with the normalized uncertainty parameter set is derived by combining the above
two LFTs as a Redheffer star product [17, 115].
Theorem 4.4 Suppose a real matrix described by an extended matrix polytope as

l
ΔX = νl+1 νi X i ∈ Rn×m (4.16)
i=1
where l is a finite natural number with 1 ≤ l < ∞, X i ’s are constant vertex matri-
l uncertainty parameters νi ’s have dependent constraints that 0 ≤ νi ≤ 1 and
ces,
i=1 νi = 1 for i = 1, . . . , l, and that |νl+1 | ≤ 1. Then, the extended matrix-
polytope-based model can be represented in an LFT form as follows:
ΔX = Fu (S(Ψ2 , Ψ1 ), Δ) (4.17)
⎡ ⎤
0 0 0 . . . 0 X1 − X2
⎢ In 0 0 . . . 0 X 2 − X 3 ⎥
⎢ ⎥
⎢ 0 In 0 . . . 0 X 3 − X 4 ⎥
⎢ ⎥
Ψ1 = ⎢ .. . . . . . . .. .. ⎥ ∈ Rn(l+1)×(nl+m) (4.18)
⎢ . . . . . . ⎥
⎢ ⎥
⎣ 0 . . . 0 In 0 Xl ⎦
0 . . . 0 0 In 0
⎡ ⎤
0 Inl
Ψ2 = ⎣ 21 In(l−1) 0 21 In(l−1) 0 ⎦ ∈ R2nl×2nl (4.19)
0 In 0 0
Δ = blockdiag[δ1 In , . . . , δl In ] (4.20)
where δi ’s are normalized real perturbational parameters with |δi | ≤ 1, and S(·, ·)
denotes a Redheffer star product.
Proof First, we prove that this representation with the dependent constraints in (4.16)
is transformed into the following with the independent constraints with respect to
the uncertainty parameter ξi ’s:

l
l−1
ΔX = ξl {(1 − ξk−1 ) ξ j }X k (4.21)
k=1 j=k
where ξ ’s have independent constraints that 0 ≤ ξ j ≤ 1 for j = 1, . . . , l −1, ξ0 = 0,

and ξl = νl+1 . To prove that the representations (4.16) and (4.21) are equivalent, it
suffices to prove that the following l-tuple parameter sets ν and ρ coincide.

l
ν := {(ν1 , . . . , νl ) : 0 ≤ νi ≤ 1, νi = 1}
i=1
l−1
(ρ1 , . . . , ρl ) : ρi = (1 − ξi−1 ) ξ j , 0 ≤ ξi ≤ 1
ρ := j=i (4.22)
for i = 1, . . . , l − 1, ξ0 = 0
By applying some algebraic manipulations to ρ, it is immediately obtained that

l
i=1 ρi = 1 and 0 ≤ ρi ≤ 1; thus ρ ⊂ ν. Then, we consider the converse. Let
ν 0 = (ν01 , . . . , ν0l ) ∈ ν with
0 ≤ ν0i < 1, and then there exist the unique ξ0i ’s
such that ν0i = (1 − ξ0(i−1) ) l−1 j=i ξ0 j , which is obtained in the following recursive
manner:
ξ0(l−1) = 1 − ν0l
ξ0(l−2) = 1 − ξ0(l−1)
1
ν0(l−1)
..
.
ξ01 = 1 − ξ0(l−1) ξ0(l−2)
1
···ξ02 ν02
and 0 < ξ0i ≤ 1 is easily confirmed. Next, let us consider the case where ν0s =
1 (1 ≤ s ≤ l) and ν0i = 0 (i
= s). In this case, by setting ξ0(s−1) = 0 and ξ0i =

1 (i
= s−1, 0), ν0i can be represented by the same form ν0i = (1−ξ0(i−1) ) l−1
j=i ξ0 j .
Hence, ν = ρ is proved.
Then, we can proceed to an LFT representation of the extended matrix polytope.
In fact, (4.21) can be described by the following LFT:
ΔX = Fu (Ψ1 , Ξ )
⎡ ⎤
0 0 0 . . . 0 X1 − X2
⎢ In 0 0 . . . 0 X 2 − X 3 ⎥
⎢ ⎥
⎢ 0 In 0 . . . 0 X 3 − X 4 ⎥
⎢ ⎥
Ψ1 = ⎢ .. . . . . . . .. .. ⎥ ∈ Rn(l+1)×(nl+m)
⎢ . . . . . . ⎥
⎢ ⎥
⎣ 0 . . . 0 In 0 Xl ⎦
0 . . . 0 0 In 0
Ξ = blockdiag[ξ1 In , . . . , ξl In ], (4.23)
where I denotes the identity matrix (when it is necessary to show the dimension
n explicitly n will be added as the subscript) and 0 denotes a zero matrix with the
appropriate dimension.
Further, the perturbational matrix Ξ is transformed into the LFT with a normalized
perturbational matrix as follows:
Ξ = Fu (Ψ2 , Δ)
⎡ ⎤
0 Inl
Ψ2 = ⎣ 21 In(l−1) 0 21 In(l−1) 0 ⎦ ∈ R2nl×2nl
0 In 0 0
Δ = blockdiag[δ1 In , . . . , δl In ] (4.24)
where δi ’s are normalized real perturbational parameters with |δi | ≤ 1.

Consequently, by combining the above two LFTs in (4.23) and (4.20) as the star
product, the final formula is obtained:
ΔX = Fu (Ψ1 , Fu (Ψ2 , Δ))

= Fu (S(Ψ2 , Ψ1 ), Δ). (4.25)
Thus, the proof is completed.

Furthermore, Theorem 4.4 can be also applied to represent a conventional matrix

polytope in an LFT form as presented in the following corollary:
Corollary 4.1 Suppose a real matrix described by a conventional matrix polytope as

l
ΔX = λi X i ∈ Rn×m (4.26)
i=1
wherel is the number of the constant vertex matrices X i ’s, 0 ≤ λi ≤ 1 (i = 1, . . . , l)

and li=1 λi ≤ 1. Then, the matrix polytope can be represented in an LFT form as
follows:
ΔX = Fu (S(Ψ2 , Ψ1 ), Δ), (4.27)

0 Inl
Ψ2 = 1 1 ∈ R2nl×2nl, (4.28)
2 I nl 2 I nl
Δ = blockdiag[δ1 In , . . . , δl In ], (4.29)
where Ψ1 is the same as in (4.18), δi ’s are normalized real perturbational parameters

with |δi | ≤ 1.
Proof First, ΔX in (4.26) can be rewritten as:

l
ΔX = νl+1 νi X i ∈ Rn×m , (4.30)
i=1

where 0 ≤ νi ≤ 1, li=1 νi = 1 for i = 1, . . . , l + 1. By noting that the range of
νl+1 is not [−1, 1] but [0, 1] as νi (i
= l + 1), Theorem 4.4 can be applied with a
modification of Ψ2 as:

0 Inl
Ψ2 = 1 1 ∈ R2nl×2nl, (4.31)
2 Inl 2 Inl
which yields (4.27). Thus, the proof is completed.
4.4 Control Design and Analysis
This section presents our proposed control design method exploiting H∞ control to
solve the problem defined in Chap. 2. The control design procedure consists of several
parts, which are introduced in the sequel as follows. First, a nonlinear state-feedback
scheme to reduce the nonlinearity of the system is introduced and then the resulting
system is defined as a virtual linear plant. Second, the frequency-dependent control
objectives and a sensitivity function shaping strategy using linear state-feedback

control to achieve those objectives are presented, which is central to dealing with
disturbances due to the base oscillation. Next, the uncertainty representation via
an extended matrix polytope for the inertia matrix is discussed with some detail
using the illustrative model and physical parameters in Table 3.4. After the preceding
procedures, the problem is finally reduced to the framework of H∞ control.
4.4.1 Nonlinear State Feedback and Virtual Linear Plant
To reduce the nonlinearity of the dynamical model of OBMs in (2.1) except the base
oscillation induced disturbance H (q, q̇, q̇ b , q̈ b ) and to construct a virtual linear
plant, the following typical nonlinear state-feedback scheme based on the nominal
parameters is employed.
τ = Cn (q, q̇)q̇ + G n (q, q b ) + D q̇ + Mn (q)u (4.32)
where the terms with the subscript (·)n denote the corresponding terms with the
nominal parameters in (2.1) as introduced in Sect. 3.3, and u is a new control torque
vector.
Substituting (4.32) into the model (2.1) and then pre-multiplying by Mn−1 (q)
yield the following dynamical system, which is defined as a virtual linear plant
with the time-varying perturbation of the inertia matrix ΔI (q) and the input-port
disturbance d.
(I + ΔI (q))q̈ = u + d, (4.33)
where
d = −Mn−1 (q){ΔC(q, q̇)q̇ + ΔG(q, q b ) + H (q, q̇, q̇ b , q̈ b )}, (4.34)
ΔI (q) = Mn−1 (q)(M(q) − Mn (q))

ΔI11 ΔI12
= , (4.35)
ΔI21 ΔI22
ΔC(q, q̇) = C(q, q̇) − Cn (q, q̇), (4.36)
ΔG(q, q b ) = G(q, q b ) − G n (q, q b ). (4.37)
4.4 Control Design and Analysis 35
The key points of the virtual linear plant in (4.33) are:

• in order to apply linear H∞ control, the system is essentially regarded as a linear
system;
• with respect to model uncertainties, the uncertainty of the inertia matrix, which is
the most influential to the overall system, is explicitly modeled, while the other
uncertainties in C(q, q̇) and G(q, q b ) are absorbed into the disturbance d together
with the base-oscillation induced disturbance H (q, q̇, q̇ b , q̈ b );
• excluding d, the system from u to q are completely decoupled and consists of
the same single-input single-output (SISO) systems, i.e., double integrators, in the
nominal case.
It should be noted that the effects of ΔC(q, q̇) and ΔG(q, q b ) in d are much less
than that of H (q, q̇, q̇ b , q̈ b ) and can be attenuated by applying linear state-feedback
control as explained later, however, H (q, q̇, q̇ b , q̈ b ) is not the case and to be overcome
by an H∞ control scheme. Further, since the system can be considered to consist of
the same SISO systems, the control design procedure in the sequel can be applied to
each SISO system, which will make the control design method very simple regardless
of the system order. Hence, this feature is one of the remarkable advantages of this
method.
Here, we discuss the gravity compensator G n (q, q b ) in (4.32). In the case of local
coordinate problems, measurements of the base oscillation are not used to generate
the reference trajectories to be tracked by the manipulator, and are used only for the
gravity compensator G n (q, q b ). Therefore, if the control system can work without
the measurements, that is, any oscillation sensors are not necessary, the system can
be constructed with lower cost, which will be practically advantageous.
Then, let us consider the gravity compensator without the information of q b . For
the illustrative model, its specific form is as follows:
G l (q) = [G l1 , G l2 ]T ,
G l1 = −{m 1 a1 sin(q1 ) + m 2n (l1 sin(q1 ) + a2n sin(q1 + q2 ))}g,
G l2 = −m 2n a2n sin(q1 + q2 )g. (4.38)
The subscript (·)l expresses that the compensator G(q) is for local coordinate prob-
lems. Note that seemingly G l (q) does not exactly match G(q, q b ) in (2.5), even in
the nominal case, because of the lack of the measurement of qb . Thus, to assess its
effectiveness, the quantities defined by Eqs. (4.39) and (4.40) are examined, where
the base oscillation qb is represented by a single sinusoidal wave in (2.7).
2π
ω1 ω1
α1 := {sin(A1 sin(ω1 t) + q1 ) − sin(q1 )}2 dt (4.39)
2π 0
2π
ω1 ω1
α2 := sin2 (A1 sin(ω1 t) + q1 ) dt (4.40)
2π 0
Fig. 4.6 Effectiveness of the 1

gravitational compensator
G n in the case of A1 = 15◦ ;
Power of gravitational torques

solid line α1 , dashed line α2 0.8 α2
0.6
0.4
0.2
α1
0
−150 −100 −50 0 50 100 150
q1 (° )
where α1 is associated with the power of the gravitational torque with the com-
pensation, while α2 without it. These values are invariant with respect to the base
oscillation angular frequency ω1 and can be interpreted as that if α2 is lager than
α1 , then it implies that the compensation scheme is effective. Then some numerical
examples with A1 = 5, 10 and 15◦ were examined for each q1 and the case of 15◦
is shown in Fig. 4.6. As seen in the figure, α2 is larger than α1 over the range of
q1 . The results of the other cases of 5 and 10◦ were similar to that and even better.
Therefore, we have confirmed that the gravitational compensation can successfully
suppress the gravitational torque. In fact, it can be demonstrated by simulations that
the control system works effectively without measurements of the base oscillation
for local coordinate problems.
4.4.2 Extended Matrix-Polytope-Based Model Uncertainty

Representation for the Inertia Matrix
Here we present the way of representing model uncertainties in an LFT form for
ΔI (q) in (4.35) of the virtual linear plant in (4.33). By employing the parameter
variation data for the illustrative model shown in Table 3.4, we analyze how ΔI (q)
varies according to the payload variation and q. Then, the idea of extended matrix
polytope in Sect. 4.3 will be applied to ΔI (q).
Recall the analysis on the parameter variation due to the payload change in
Sect. 3.3. The inertia matrix variation ΔM(q) = M(q) − Mn (q) in (3.7) is linear in
Δm 2 . Thus, substituting (3.7) into (4.35) yields
ΔI (q) = Mn−1 (q)ΔM(q)

= Δm 2 Mn−1 (q)M (q), (4.41)
ΔI
11
0.05
−0.05
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
ΔI
12
0.02
−0.02
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
ΔI
21
1
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
ΔI
22
0.5
−0.5
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
cos(q )
2
Fig. 4.7 ΔIij ’s versus cos(q2 ) with various payloads. The solid lines represents the cases of
Δm 2 > 0, and the dashed one do the case of Δm 2 < 0
which shows that ΔI (q) is also linear in Δm 2 . For all the parameter cases in Table 3.4,
that is, the payload widths 0, 1, 2, . . ., 10 mm, each element ΔIij (i, j = 1, 2) in (4.41)
is calculated over q2 ∈ [0, 180]◦ , and the results of which are shown in Figs. 4.7, 4.8
and 4.9 respectively. Note that ΔIij (i, j = 1, 2) is an even function of q2 and, thus,
the calculation over nonnegative q2 suffices.
Figure 4.7 shows that ΔIij ’s versus cos(q2 ) on the different scales on the respective
vertical axes, where the solid lines represent the cases of positive variation of payload
mass (Δm 2 > 0) and the dashed ones do the negative cases (Δm 2 < 0). As seen
from the figure, ΔIij ’s, except ΔI11 , are almost linearly dependent on cos(q2 ). Then,
see Fig. 4.8 which displays the same data as in Fig. 4.7 on a uniform vertical axis
scale. It is seen that ΔI11 and ΔI12 are much less than ΔI21 and ΔI22 . If we regard
the set of ΔIij ’s as a set of four-dimensional vector in R4 , the set can be a subset of
two-dimensional manifold with coordinates (Δm 2 , cos(q2 )). Furthermore, the facts
that ΔIij ’s are completely linear in Δm 2 and almost linear in cos(q2 ) considering
that ΔI11 is much less than the other ΔIij ’s suggest that we can approximate the set
of ΔIij by a subset of a hyperplane in R4 . To understand this claim intuitively, three-
dimensional graphs of ΔIij ’s are presented in Fig. 4.9, where the set is very similar to
Area 3 in Fig. 4.5. Thus, the set of ΔI (q) can be represented by an extended matrix
polytope with merely two vertex matrices as in (4.15) and subsequently by its LFT
form. And see Table 4.1 that presents the minimum and maximum values of ΔIij .
If one takes into account only the information in Table 4.1 to represent the model
ΔI
11
1
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
ΔI
12
1
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
ΔI21
1
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
ΔI
22
1
−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
cos(q )
2
Fig. 4.8 Comparison of ΔIij ’s on the uniform scale
0.6 0.6
0.4 0.4
width: 10 mm (q2=180°) width: 10 mm (q2=180°)
0.2 0.2
0 0
11
12
ΔI
ΔI
−0.2 −0.2
−0.4 −0.4
width: 0 mm (q2=0°) width: 0 mm (q2=0°)
0.5 0.5
0.5 0.5
0 0
0 0
ΔI −0.5 −0.5 ΔI21 ΔI −0.5 −0.5 ΔI21

22 22
Fig. 4.9 3D representation of the ΔIij ’s space

Table 4.1 Range of ΔIij variation

ΔIij Min Max
ΔI11 −0.0393 0.0393
ΔI12 −0.0141 0.0141
ΔI21 −0.5591 0.5591
ΔI22 −0.2740 0.2740
uncertainty, the representation is based on the conventional matrix polytope resulting

a hyper-polyhedron with 16 vertices (recall Area 1 in Fig. 4.5), which is obviously
too conservative. Therefore, The above argument justifies that the approach based
on extended matrix polytopes is effective and significant.
Next, we analyze in a quantitative manner how the set of ΔIij ’s (two-dimensional
manifold) can be approximated by a hyperplane modeled with the following extended
matrix polytope:
ΔI (q) ≈ ν3 (ν1 ΔI1 + ν2 ΔI2 ), (4.42)

0.0336 0.0141
ΔI1 = , (4.43)
0.5591 0.2454

0.0054 −0.0141
ΔI2 = , (4.44)
−0.0790 0.2736
where ΔI1 and ΔI2 are vertex matrices respectively corresponding to the cases of
q2 = 0◦ and q2 = 180◦ with the same payload width of 10 mm, νi ∈ R, 0 ≤ νi ≤ 1
2
for i = 1, 2, i=1 νi = 1, and |ν3 | ≤ 1. ν1 and ν2 are associated with cos(q2 (t)), and
ν3 with Δm 2 . Thus, once the payload has been fixed, ν3 is time-invariant, whereas
ν1 and ν2 is time-varying during the manipulator motion.
In order to facilitate the analysis, let us consider a bijective mapping from R2×2
to R4 and its inverse defined as follows. Let X ∈ R2×2 be

X 11 X 12
X= , (4.45)
X 21 X 22
then, the mapping vec(·) and its inverse vec−1 (·) are defined as:
vec(X ) := [X 11 , X 12 , X 21 , X 22 ]T ∈ R4 , (4.46)
vec−1 (vec(X )) := X. (4.47)
Hence, using the tools, we analyze sort of feasibility of the approximation of

ΔI (q) in (4.41) by (4.42) quantitatively. Instead of the set of ΔI (q)’s variation, we
consider the set of vec(ΔI (q))’ variation in R4 which is denoted by M. Next, let M∗
be the hyperplane spanned by vec(ΔI1 ) and vec(ΔI2 ). Let us define a orthogonal
projection of d I ∈ M onto M∗ being denoted by d I ∗ = P(d I) ∈ M∗ (see
Fig. 4.10 Projection dI

onto M∗ M
de
dI 1
dI*
M*
dI 2
O
Fig. 4.10). Since the approximation error vector de = d I − d I ∗ is orthogonal to

M∗ , i.e., vec(ΔI1 ) and vec(ΔI2 ), d I ∗ for given d I can be obtained uniquely and
straightforwardly as:
−1
< d I1 · d I1 > < d I1 · d I2 > < d I1 · d I >
d I ∗ = [d I 1 , d I 2 ] ,
< d I2 · d I1 > < d I2 · d I2 > < d I2 · d I >
(4.48)
where d I 1 = vec(ΔI1 ), d I 2 = vec(ΔI2 ), and < · > denotes Euclidean inner

product defined in R4 .
Being based on (4.48), we calculate the following criterion to evaluate the approx-
imation:
max d I∈M d I − d I ∗
1 = , (4.49)
max d I∈M d I|
where · denotes the Euclidean norm. 1 indicates how close to a hyperplane

geometrically M is in R4 , that is, the smaller 1 implies the better approximation.
In this case, the resulting was 0.0279, which suggests that the approximation is
fairly good.
Moreover, using the inverse mapping vec−1 (·), we evaluate the feasibility of the
approximation in R2×2 . Specifically, the following criterion is computed:
2 = max σ̄ (vec−1 (d I − d I ∗ )), (4.50)

d I∈M
where σ̄ (·) denotes the largest singular value. And the resulting 2 was 0.0170. ΔI (q)
is a displacement matrix from the identity matrix I2 , hence the approximation error
criterion 2 should be evaluated relatively to 1. In this sense, we can conclude that
the approximation is fairly good.
Remark 4.3 The above argument is basically based on the viewpoint that how less
conservative and simple model uncertainty representation can be for control design
using the H∞ control framework. For stringently evaluating the feasibility of the
approximation, the robust stability of the resulting control system based on the
uncertainty modeled by (4.42) should be analyzed over vec−1 (M). However, at

this point, the effectiveness and significance of the developed machinery, extended
matrix polytopes, has been satisfactorily demonstrated.
Finally in this subsection, we derive an LFT form for ΔI (q) approximated by
(4.42). Applying Theorem 4.4 to ΔI (q) in (4.42) yields the following LFT represen-
tation:
ΔI (q) = Fu (S(Ψ2 , Ψ1 ), Δ I ), (4.51)

⎡ ⎤
0 0 ΔI1 − ΔI2
Ψ1 = ⎣ I2 0 ΔI2 ⎦, (4.52)
0 I2 0
⎡ ⎤
0 I4
Ψ2 = ⎣ 21 I2 0 21 I2 0 ⎦ , (4.53)
0 I2 0 0
⎡ ⎤
0 0 ΔI1 − ΔI2
S(Ψ2 , Ψ1 ) = ⎣ 21 I2 0 21 (ΔI1 + ΔI2 ) ⎦ , (4.54)
0 I2 0
I := {blockdiag[δ1 (t)I2 , δ2 I2 ] : δi ∈R, |δi | ≤ 1}, (4.55)
ΔI ∈ I , (4.56)
where δ1 (t) is associated with cos(q2 (t)), thus time-varying, whereas δ2 is time-
invariant correspondingly to Δm 2 . Further, for the control design, the following
equation will be utilized:
I + ΔI (q) = Fu (Φ, Δ I ), (4.57a)

Φ11 Φ12
Φ= (4.57b)
Φ21 Φ22
⎡ ⎤
0 0 ΔI1 − ΔI2
= ⎣ 21 I2 0 21 (ΔI1 + ΔI2 ) ⎦ . (4.57c)
0 I2 I2
Note that the above notion and methodology is obviously not exclusive for the
OBM problems and extensively available for general robotic manipulator control
problems.
4.4.3 Sensitivity Function Shaping Strategy Using Linear

State-Feedback Control
In order to solve the oscillatory base problem, it is crucial to cope with disturbances
due to the base oscillation, which is exerted at the input port of the plant. In this
Fig. 4.11 Diagram of a d n

closed-loop system
r + e u+ y +
_ K P
+ +
work, this issue is captured as a filtering problem, i.e., how to make the sensitivity of
the system to the disturbance as low as possible. Hence, the issue is, specifically, to
construct desirable frequency-dependent properties of the corresponding sensitivity
functions of the system. In this section, we present a strategy for shaping sensitivity
functions using linear state-feedback control for H∞ control design.
By being based on the virtual linear plant in the nominal case, we consider each
decoupled SISO system, that is, from u i to qi (i = 1, 2), whose state-space and
output equations are as in the following:

01 0
ẋ = x+ u, (4.58a)
00 1 i

yi = 1 0 x, (4.58b)
where x = [qi , q̇i ]T . Then, Fig. 4.11 depicts a closed-loop system, where P and K
represent the transfer functions of this SISO system and a controller to be designed
respectively; ri is the reference trajectory to be tracked by yi ; n is a sensor noise;
ei = ri − yi − n is the tracking error with noise; and di is a disturbance at the input
port. Then the following sensitivity functions are defined as standard [115].
S := (1 + P K )−1 (ri → ei , n → −ei ) (4.59)

S P := (1 + P K )−1 P (di → −ei ) (4.60)
Ta := K (1 + P K )−1 (ri → u i , n → −u i ) (4.61)
S is referred to as sensitivity function which represents tracking control performance

and noise sensitivity; pre-multiplying S by the plant P yields S P , which pertains to
the input port disturbance attenuation property and, hence, plays a key role in coping
with disturbances due to the base oscillation; Ta is referred to as quasi-complementary
sensitivity function. Ta is the sensitivity from ri and n to the control input, and will
be utilized not only when in need of suppressing the control input, but also when
regarding an additive unmodeled dynamics. Furthermore, Ta is necessary to be taken
into account so as to make the H∞ control design problem the standard problem
[16, 115].
Before introducing the method of shaping the above functions, the frequency
domain characteristics of the disturbance d in (4.33) is evaluated. In particular, torque
H in (2.6) is focused on, which will be dominant in d. Assume the steady-state
Fig. 4.12 Desirable

frequency responses of the
sensitivity functions
S Sp T
Gain
anti-ω mode
Frequency
manipulator operating frequency ωs and examine Eqs. (2.6) and (2.7), it is then
found out that the principal frequency of H is expected to be ωi and that 2ωi ,
ωs + ωi and |ωs − ωi | will also appear since there exist the products of q˙b , q̇1 , and
q̇2 . Then, the minimal frequency range Ω such that contains all those frequencies
can be determined. Thus, the discussion is summarized as “to attenuate the effect of
d it is necessary that S P should be low enough over the frequency range Ω.”
Now let us move on to the sensitivity function shaping strategy. Taking the above
necessary S P property into account together with the standard requirements for S and
T , the desirable frequency responses of those functions are as depicted in Fig. 4.12.
Focus on the anti-resonant modes of S and S P , labeled with “anti-ω modes" in the
figure, which will satisfy the above requirement for S P . As seen from (4.59) and
(4.60), once the plant is determined, S and S P cannot be independently constructed
by designing K . Therefore, the idea is to achieve this frequency-dependent objective
for S P by shaping S instead of S P , which gives rise to anti-ω modes to S. It should
be noted that the anti-ω modes for S is necessary for global coordinate problems
where the reference trajectory necessarily contains ωi ’s modes to cancel out the base
oscillation while it is not the case for local coordinate ones. On the other hand, recall
that the plant in (4.58) has a double integrator, which increases the gains of S P at
low frequencies as compared to those of S and then will degrade the disturbance
attenuation performance. Therefore, to maintain the relation between S and S P at
low frequencies, a linear state-feedback control F x = [F1 , F2 ]x is added to the plant
so that P has a flat gain property with less than zero dB at low frequencies and hence
the requirements are achieved. Thus, the model in (4.58) is rewritten as in (4.62):

0 1 0
ẋ = x+ u, (4.62a)
−F1 −F2 1 i

yi = 1 0 x, (4.62b)
where u i is the new control input. F1 which controls the relative gain of S P to S should
be chosen subject to the requirement for S P , then F2 can be determined such that with
respect to the F1 the nominal model in (4.62) has an allowable damping property.
Theoretically, the larger F1 gain will lead to the smaller gain of S p . However, in the
presence of sensor error and actuator saturation, F1 cannot be made arbitrarily large,
and neither can F2 . Therefore, an optimization scheme is applied to determine the
gain F, which is based on control simulations and numerical direct search.
For given F, a controller is synthesized and a step tracking control simulation
is conducted as mentioned later. Then, the control error vector e is calculated. By
extracting steady-state elements from e, e s is obtained, and its Frobenius norm ||e s || F
is set to the criterion to be minimized. Since the H∞ control synthesis is via a
suboptimal scheme and the obtained ||e s || F is thus not continuous in F, nonlinear
optimization schemes using the derivatives cannot be used. Instead, a direct search
method was employed over F1 × F2 = [20, 21, . . . , 89, 90] × [20, 21, . . . , 89, 90],
which range had been determined in advance through a trial and error manner via
simulations. The resultant F is [66, 24].
Consequently, the proposed strategy involving the linear state-feedback control
scheme is reduced to a mixed-sensitivity problem for S and Ta for the plant in (4.62),
where S p will be indirectly shaped through shaping S respectively. Note that the
linear state-feedback control scheme to shape S P distinguishes our methodology
from the conventional ones, e.g., [91], which employ a PD control to stabilize the
double integrator after linearization.
4.4.4 Generalized Plant for H∞ Control Design and Analysis
We discuss here construction of generalized plants for H∞ control design and analy-
sis. To demonstrate and evaluate the extended polytope-based model uncertainty
representation and the weighting function matrix Wn for the complementary sensi-
tivity function Ta , we consider four types of system configurations as depicted in
Figs. 4.13 and 4.14, where P and K denote the two-input two-output plant and con-
troller, respectively, although the same symbols denote SISO systems in Fig. 4.11.
Ws , Wu are also the weighting function matrices associated with the sensitivity func-
tion S and the quasi-complementary sensitivity function Ta respectively. In Fig. 4.13,
at the top, System configuration 1 contains neither model uncertainty nor Wn ; the
second one 1b does not contain Wn but the model uncertainty ΔI ; in Fig. 4.14 Sys-
tem configuration 2 does not contain model uncertainty but Wn ; and the last one 2b
contains both the model uncertainty and Wn .
Ws for S is defined depending on the base oscillation models as follows:

ws 0
Ws = , (4.63)
0 ws
0.4π
nm
s 2 + 2ωi s + ωi2
ws = , (4.64)
s + 10−5 s 2 + 0.02ωi s + ωi2
i=1
Fig. 4.13 System

configurations for H∞
control design and analysis Wu -r
without Wn
u y + -e
P Ws
+
System configuration 1
Wu -r
u y + -e
P(ΔI) +
Ws
System configuration 1b
Fig. 4.14 System n

configurations for H∞
control design and analysis
with Wn
Wu Wn -r
u + -e
P Ws
y +
System configuration 2
n
Wu Wn -r
u + -e
P(ΔI) +
Ws
y
System configuration 2b
2
10
single
double
Bretschneider
1
10
0
10
Gain
−1
10
−2
10
−3
10 −1 0 1 2 3
10 10 10 10 10
Frequency (rad/s)
Fig. 4.15 Frequency responses of ws ’s
where ws includes a notch-type filter to let S p and S have anti ωi modes corre-
sponding to the base oscillation, and a quasi-integrator for step tracking. For the
single-frequency oscillation, n m = 1, ω1 = 2π rad/s; for the double-frequency
oscillation, n m = 2, ω1 = π , ω2 = 2π rad/s; and for the Bretschneider oscilla-
tion, the setting as n m = 2, ω1 = 1.1π , ω2 = 1.7π rad/s works well regardless
of n ω = 10, which is because the Bretschneider frequencies are very close and
have a single-peak frequency as shown in Fig. 2.6. For a small range of oscillation
frequency variation, the robustness can be ensured by using multiple ωi weights as
above. However, for a large range of variation, e.g., a case of a vessel with a variety
of traveling velocities, some adaptive schemes would be required, which is one of
the problems to be tackled in the future. The respective frequency responses of ws ’s
are shown in Fig. 4.15. Then, Wu and Wn are set as follows:

wu 0
Wu = , (4.65)
0 wu
10−4 s + 10−2
wu = , (4.66)
s + 103

wn 0
Wn = , (4.67)
0 wn
20(s + 10π )
wn = . (4.68)
s + 37π
2
10
w
n
w
1 u
10
0
10
−1
10
Gain
−2
10
−3
10
−4
10
−5
10 −1 0 1 2 3
10 10 10 10 10
Frequency (rad/s)
Fig. 4.16 Frequency responses of wu and wn
where wu is a high-pass filter for the control input u and wn is a high-pass filter
for sensor noises and additive model uncertainties. As shown in Fig. 4.16, wu is
nonrestrictive compared with wn .
Based on the respective system configurations, corresponding generalized plants
are constructed. The generalized plants are utilized in H∞ controller synthesis and
robustness analysis. Controller synthesis methods adopted depends on the system
configurations; in the case of System configurations 1 and 2 in Figs. 4.13 and 4.14
which do not contain the model uncertainty ΔI , since the plant is decoupled, “γ -
iteration” for a SISO H∞ controller is conducted; whereas in the case of the con-
figurations 1b and 2b with ΔI , “D-K iteration” for an MIMO H∞ controller is
performed.
First, we consider the case of System configurations 1 and 2, where every block
consists of decoupled identical SISO systems. Then, we can deal with the system
as a SISO system within these frameworks. Hence, the state-space representation of
P is given by (4.62). Augmenting P by additional inputs and outputs, the reference
signal ri , the control error ei of Link i, for performance evaluation, yields

0 1 0 0
ẋ = x+ (−ri ) + u, (4.69a)
−F1 −F2 1 1 i

−ei 10 1 0
= x+ (−ri ) + u, (4.69b)
ui 00 0 1 i

yi = 1 0 x − ri , (4.69c)
which does not include sensor noises, and is denoted by Σ1 for System configuration
1. Further, multiplying Σ1 with the weighting matrix
W1 = diag[ws , wu , 1] (4.70)
yields the generalized plant Paug1 = W1 Σ1 .

For System configuration 2, Σ2 is defined as:

0 1 00 −ri 0
ẋ = x+ + u, (4.71a)
−F1 −F2 10 ni 1 i

−e 10 11 −ri 0
= x+ + u, (4.71b)
ui 00 00 ni 1 i

−ri
yi = 1 0 x + 1 1 , (4.71c)
ni
where the sensor noise n i is taken into account. The weighting function matrices are
defined as
W2l = diag[ws , wu , 1], (4.72)

W2r = diag[1, wn , 1], (4.73)
with which the generalized plant is constructed as Paug2 = W2l Σ2 W2r .

Paug1 and Paug2 are employed to synthesize corresponding H∞ controllers via
γ -iteration.
Next, we address the case of System configurations 1b and 2b with an explicit
model uncertainty representation. Figure 4.17 depicts the generalized plant with the
dashed line box for System configuration 2b in Fig. 4.14, and the controller to be
synthesized K , the normalized structured model uncertainty matrix Δ I in (4.55), and
additionally the fictitious full-block complex uncertainty matrix Δ p with σ̄ (Δ p ) ≤ 1
intended to accommodate robust performance.
To obtain an LFT representation of P(ΔI ) for System configurations 1b and 2b
as in Fig. 4.17, applying the formula of the inverse of a LFT, e.g., [115], to the LFT
of I + ΔI in (4.57) yields
Fu (Φ , Δ I ) = (I + ΔI )−1
= Fu (Φ, Δ I )−1 , (4.74a)

Φ11 Φ12
Φ = Φ
Φ21 22
−1 −1
Φ11 − Φ12 Φ22 Φ21 −Φ12 Φ22
= −1 −1
Φ22 Φ21 Φ22
⎡ ⎤
0 −(ΔI1 − ΔI2 ) −(ΔI1 − ΔI2 )
= ⎣ 21 I2 − 21 (ΔI1 + ΔI2 ) − 21 (ΔI1 + ΔI2 ) ⎦ . (4.74b)
0 I2 I2
Fig. 4.17 Generalized plant

for H∞ controller synthesis
Δ
and robustness analysis
based on System Δp
configuration 2b
ΔI
n -r
Wu Wn
u P’ + -e
+
Ws
y
Paug
K
Thus, P(ΔI ) = Fu (P , Δ I ) such that the state-space representation of P is

expressed by

0 I2 0 0
ẋ = −F Φ x + w + u, (4.75a)
−F1 Φ22 2 22 Φ21 Φ22
−F Φ

z = −F1 Φ12 2 12 x + Φ11 w + Φ12 u, (4.75b)

y = I2 0 x, (4.75c)
where x = [q1 , q2 , q̇1 , q̇2 ]T , y = [q1 , q2 ]T , u = [u 1 , u 2 ]T , the new control input

defined in (4.32), z ∈ R4 , w ∈ R4 . With a slight abuse of notation, z and w are
related as w = Δ I z.
Further, by incorporating additional inputs and outputs similarly to Paug1 and
Paug2 , i.e., the reference vector −r, the sensor noise vector n, and the control
error vector e, the augmented plant for System configuration 1b denoted by Σ1b
is expressed as follows:

0 I2 0 0 w 0
ẋ = −F Φ x + 0 + u, (4.76a)
−F1 Φ22 2 22 Φ21 −r Φ22
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
z −F Φ
−F1 Φ12 2 12 Φ11 0 Φ12
⎣ −e ⎦ = ⎣ ⎦ x + ⎣ 0 I2 ⎦ w
I2 0 + ⎣ 0 ⎦ u, (4.76b)
−r
u 0 0 0 0 I2

w
y = I2 0 x + 0 I2 . (4.76c)
−r
Multiplying Σ1b by the weighting function matrix
W1b = blockdiag[I4 , Ws , Wu ], (4.77)
the generalized plant Paug1b = W1b Σ1b can be obtained.

Moreover, by taking into account n, the augmented plant Σ2b is given by
⎡ ⎤
w
0 I2 0 00 ⎣ ⎦ 0
ẋ = −F Φ x + 0 0 −r + u, (4.78a)
−F1 Φ22 2 22 Φ21 Φ22
n
⎡ ⎤ ⎡ −F Φ
⎤ ⎡ ⎤⎡ ⎤ ⎡ ⎤
z −F1 Φ12 2 12 Φ11 0 0 w Φ12
⎣ −e ⎦ = ⎣ I2 0 ⎦ x + ⎣ 0 I2 I2 ⎦ ⎣ −r ⎦ + ⎣ 0 ⎦ u,
u 0 0 0 0 0 n I2
(4.78b)
⎡ ⎤
w
y = I2 0 x + 0 I2 I2 ⎣ −r ⎦ . (4.78c)
n
Then, defining weighting function matrices as
W2bl = blockdiag[I4 , Ws , Wu ], (4.79)

W2br = blockdiag[I6 , Wn ], (4.80)
the generalized plant for System configuration 2b is acquired as Paug2b = W2bl

Σ2b W2br .
Consequently, being based on the obtained generalized plants Paug1b or Paug2b ,
and employing the corresponding structured uncertainty matrix
Δ = blockdiag[Δ I , Δ p ], (4.81)
D–K iteration is performed to synthesize an H∞ controller. Depending on the gen-

eralized plants, the dimension of Δ is defined as in the following:
1 := {blockdiag[δ1 (t)I2 , δ2 I2 , Δ P ] :
δi ∈ R, |δi | ≤ 1, Δ P ∈ C2×4 , σ̄ (Δ P ) ≤ 1}. (4.82)
for Paug1b , and
2 := {blockdiag[δ1 (t)I2 , δ2 I2 , Δ P ] :
δi ∈ R, |δi | ≤ 1, Δ P ∈ C4×4 , σ̄ (Δ P ) ≤ 1}. (4.83)
for Paug2b respectively, where 1 and 2 are the uncertainty sets which possible Δ
belongs to.
4.4.5 H∞ Controller Synthesis and Robustness Analysis
4.4.5.1 H∞ Controller Synthesis
Now we demonstrate H∞ controller synthesis and robustness analysis using the

constructed generalized plants. As mentioned above, being based on Paug1 or Paug2
without explicit uncertainty model, γ -iteration is conducted to obtain an SISO H∞
controller, whereas for Paug1b or Paug2b with consideration of the explicit uncertainty
model, D–K iteration is done getting an MIMO H∞ one. For each generalized plant
Paugi (i = 1, 1b, 2, 2b), three controllers denoted by K i1 , K i2 , K i3 are synthesized
according to the respective oscillation models, the single-frequency one, the double-
frequency one, the Bretschneider one.
Then, γ -iteration with Paugi (i = 1, 2) has given the resultant H∞ controller
denoted by K ij ( j = 1, 2, 3) with the upper bound of γ as
||Fl (Paugi , K ij )||∞ ≤ γij , (4.84)
where γij is shown in Table 4.2. These values of γij ’s being less than one imply that
all the controllers can provide better performance than the respective specified ones
in the nominal model case, and the differences between γ1 j ’s and γ2 j ’s show that the
weighting function wn is restrictive to an certain degree. Note that this γ -iteration-
based synthesis is equivalent to a synthesis to obtain a robust stabilizing controller
with an unstructured complex uncertainty.
Next, we present the results of D–K iteration using Paugi (i = 1b, 2b) and the
corresponding structured uncertainty matrices Δ ∈ 1 in (4.82) and Δ ∈ 2 in
(4.83), and this synthesis procedure means to solve the robust performance problem.
The resultant upper μ bound for Paugi with the index j representing the oscillation
model as above has been obtained as
sup ||Dl (iω)Fl (Paugi (iω), K ij (iω))Dr (iω)||∞ ≤ μij , (4.85)

ω∈[0,∞]
where ω is the frequency variable and Dl and Dr are the frequency-dependent com-
plex scaling matrices which satisfy that Dr Δ = ΔDl−1 . The dimensions and struc-
tures of Dl and Dr depend on Paugi ; specifically, the set of Dl and Dr involved for
Paug1b is as follows:
Table 4.2 Achievable upper γ bound in each SISO H∞ controller synthesis via γ -iteration; “sin-
gle,” “double,” “Bretschneider” represent the single-frequency oscillation, the double-frequency
one, and the Bretschneider one, respectively
Paug Single Double Bretschneider
Paug1 0.156 0.156 0.156
Paug2 0.703 0.859 0.859
Table 4.3 Achievable upper μ bound in each MIMO H∞ controller synthesis via D–K itera-
tion; “single”, “double”, “Bretschneider” represent the single-frequency oscillation, the double-
frequency one, and the Bretschneider one, respectively
Paug Single Double Bretschneider
Paug1b 0.633 0.676 0.873
Paug2b 1.601 1.479 2.019
⎧ ⎫
⎨ (Dl , Dr ) : Dl = blockdiag[D1 , D2 , d1 I4 ], ⎬
D = Dr = blockdiag[D1−1 , D2−1 , d1−1 I2 ], , (4.86)
⎩ ⎭
Di = Di∗ 0 ∈ C2×2 , d1 > 0 ∈ R
and the set for Paug2b is as follows:

⎧ ⎫
D = Dr = blockdiag[D1−1 , D2−1 , d1−1 I4 ], . (4.87)
⎩ ⎭
Di = Di∗ 0 ∈ C2×2 , d1 > 0 ∈ R
However, in practice, μij is computed over finite points logarithmically spaced in a

designated frequency range Iω = {ω1 , ω2 , . . . , ωn } as
μij = max ||Dl (iωk )Fl (Paugi (iωk ), K ij (iωk ))Dr (iωk )||∞ , (4.88)
ωk ∈Iω
where in our demonstration 100 points from 10−1 to 103 rad/s have been employed
as Iω . Table 4.3 shows the obtained μij ’s. μ1bj ’s ( j = 1, 2, 3) are less than one, which
implies that the controllers K 1bj ’s can achieve the desired robust performance without
considering Wn if Δ is time-invariant. However, since Δ is time-varying uncertainty,
strictly speaking, these upper μ bounds cannot give us definite evaluation in terms
of robust performance or robust stability either, but some reasonable properties can
be expected. Then, since Paug2b contains Wn , μ2bj ’s are greater than one, but not
that large. Similarly, the results do not present any conclusive facts either, however,
feasible controllers can be expected. At least with respect to robust stability, more
strict analysis is necessary, which will be demonstrated later.
Here it should be noted that a controller synthesized via D–K iteration, in general,
results in a considerably high-order system due to dynamical system scalings. In our
demonstration case, the order of each K ij is shown in Table 4.4, where K 1 j ’s and
K 2 j ’s are SISO controllers, while K 1bj ’s and K 2bj are MIMO ones. Since the SISO
controllers are employed as follows:
diag[K ij , K ij ], (4.89)
whose order amounts to the double of the order of K ij . As seen from the table, the
orders of K 1bj and K 2bj are too large compared to ones of the controllers obtained
Table 4.4 Orders of the obtained H∞ controllers

Controller Order Controller Order Controller Order
K 11 6 K 12 8 K 13 8
K 1b1 92 K 1b2 96 K 1b3 96
K 1b1r ed 12 K 1b2r ed 16 K 1b3r ed 16
K 21 7 K 22 9 K 23 9
K 2b1 94 K 2b2 98 K 2b3 98
K 2b1r ed 14 K 2b2r ed 18 K 2b3r ed 18
Table 4.5 Zeros and poles of the nominal plant and weighting functions
Transfer function Zeros Poles
P −2.083e+1
−3.168e+0
Ws −6.283e+0 −6.283e+0 + 6.283e+0i
−6.283e+0 −6.283e+0 − 6.283e+0i
−1.000e-5
Wu −1.000e+1 −1.000e+3
Wn −3.142e+1 −1.162e+2
via γ -iteration. Hence, we have performed model reduction to those controllers by

using the balanced realization with Hankel singular values. Each reduced order was
determined so as to be consistent with that of the corresponding controller obtained
via γ -iteration. The resultant controllers through model reduction are denoted by
adding the suffix “red,” e.g., from K 1b1 to K 1b1red . For ease of implementation and
analysis, the reduced-order controllers will be mainly used instead of the original
ones.
Then let us investigate the obtained controllers in some detail, particularly by
considering the case of single-frequency oscillation for exposition simplicity. First,
we focus on zeros and poles of the obtained controllers K i1 ’s. Table 4.5 shows zeros
and poles of the nominal plant and weighting functions, which are reflected to those
of the H∞ controllers as shown in Table 4.6. K 11 and K 21 contain all the poles of
P, Wu as zeros, and K 21 has additionally the pole of Wn as a zero. While K 1br ed22
and K 2br ed22 contain similar zeros, which however have slightly shifted from the
poles by taking into account of the model uncertainties. Considering poles of the
controllers, we see that all the controllers contain all the poles of Ws , i.e., the pole for
quasi-integrator and ones of the base oscillation frequency ω1 , as poles which have
not shifted at all even under consideration of model uncertainties. This fact represents
a considerably important feature of H∞ control which ensures a systematic control
design to implement so-called “the internal model principle” by modifying Ws . As
the above argument, in general, an H∞ controller tries to shape the objective transfer
functions by adopting zeros or poles from the poles of the given plant and weighting
functions, which can provide a control design perspective [94, 95].
Table 4.6 Zeros and poles of the obtained H∞ controllers

Controller Zeros Controller Zeros
K 11 −1.000e+3 K 21 −1.000e+3
−2.083e+1 −2.083e+1
−3.168e+0 −3.168e+0
−3.103e+1 + 3.450e+0i −3.025e+0 + 3.421+0i
−3.103e+1 − 3.450e+0i −3.025e+0 − 3.421e+0i
−1.112 e + 2
K 1b1red22 −9.409e+2 K 2bred22 −1.022e+3
−2.656e+1 −2.173e+1
−3.210e+0 −3.119e+0
−7.190e-1 + 3.411e+0i −1.979e+0 + 3.468e+0i
−7.190e-1 − 3.411e+0i −1.979e+0 − 3.468e+0i
−1.964 e+2 −1.413e+2 + 1.273e+0i
−1.413e+2 − 1.273e+0i
Controller Poles Controller Poles
K 11 −1.000e-5 K 21 −9.992e-6
−6.283e-2 + 6.283e+0i −6.283e+0 + 6.283e+0i
−6.283e-2 − 6.283e+0i −6.283e+0 − 6.283e+0i
−2.597e+1 + 4.891e+1i −3.253e+0 + 4.719e+1i
−2.597e+1 − 4.891e+1i −3.253e+0 − 4.719e+1i
−6.768e+1 −6.388e+1
−6.727e+2
K 1b1red22 −1.000e-5 K 2bred22 −1.000e-5
−6.283e-2 + 6.283e+0i −6.282e-2 + 6.283e+0i
−6.283e-2 − 6.283e+0i −6.282e-2 − 6.283e+0i
−3.071e+1 + 2.674e+1i −2.867e+1 + 3.993e+1i
−3.071e+1 − 2.674e+1i −2.867e+1 − 3.993e+1i
−4.415e+2 −2.160e+2
−6.703e+1 −7.729e+1 + 1.554e+1i
−7.729e+1 − 1.554e+1i
For the reduced-order controllers, the pole-zero cancelation pairs are omitted
Next, we present frequency responses of the controllers and the resultant sen-
sitivity functions of interest in (4.59)–(4.61). Figure 4.18 depicts those of the H∞
controllers. All the controllers have the resonant mode of ω1 and the quasi-integrator
property given by Ws , which ensures the internal model property for tracking control
and disturbance attenuation for step signals and ones of frequencies near ω1 . Com-
paring the controllers, it is seen that in the frequency range where tracking control
and disturbance attenuation are required, K 11 is the highest gain controller and K 12
is almost the same as it. However, in high frequencies of greater than 70 rad/s, the
gain of K 12 is higher than that of K 11 , which is effect of Wn restricting S in (4.59).
Addition of model uncertainties has resulted in relatively low-gain controllers so
as to attain robustness as expected, particularly K 1b1r ed is the lowest gain one, but
addition of Wn has increased the gain of K 2b1r ed also by restricting S. Consider
difference between the channels in K 1b1r ed and K 2br ed in Fig. 4.18. The diagonal
element K 1b1r edii and K 2b1r edii (i = 1, 2) are the controllers for Joint i. Comparing
the controllers for Joints 1 and 2, the gains of the controllers for Joint 1 are lower
than the corresponding ones for Joint 2, which reflects the fact that control of Joint
1 requires more robustness due to inertia parameter variation due to change of q2 .
Figure 4.19 shows the frequency responses of the resultant sensitivity functions
with the respective H∞ controllers. In addition to S, S p , and Ta in (4.59)–(4.61),
the complimentary sensitivity function T := P Ta (for multiplicative model uncer-
tainties) is depicted. The sensitivity functions with K 11 (solid line) and K 21 (dashed
line) reveal desirable profiles as depicted in Fig. 4.12. Due to the difference in the
gain profiles of those controllers, the gains of S and S p with K 21 are slightly lower
than those with K 11 in frequencies of less than 70 rad/s, while it is the opposite case
in frequencies of greater than 70 rad/s, which is difficult to be seen on a normal
scale, but it will be possible by magnifying the graph. This difference implies that, in
the nominal case, K 11 will perform slightly better tracking control and disturbance
attenuation than K 21 while K 21 will exhibit less influence of high-frequency sensor
noise than K 11 . On the other hand, T and Ta with K 11 and K 21 directly reflect the
controller gains, particularly in frequencies of greater than 70 rad/s.
Then let us see the cases of K 1b1r ed and K 2b1r ed . Similarly with the above cases,
the gain profiles of the respective sensitivity functions well reflect the properties
of the controllers. That is, since the gains of S and S p with K 1b1r ed are higher
than those with the other controllers due to consideration of model uncertainties, in
the nominal case, K 1b1r ed will exhibit poorer performance of tracking control and
disturbance attenuation. Further, the profiles with K 11 reveals unfavorable deformed
shape in frequencies near ω1 . The property of K 2b1r ed has been slightly recovered
by Wn . However, these two controllers K 1b1r ed and K 2b1r ed can be expected to
perform better than K 11 and K 21 in the worst case of model variation, which will
be demonstrated later. Moreover, we present the resultant sensitivity functions in
the cases of double-frequency oscillation and Bretschneider oscillation in Figs. 4.20
and 4.21 respectively. It should be noted that the essential properties and arguments
described above are common ones for all the frequency cases.
4.4.5.2 Robustness Analyses
Here we perform robustness analyses on the resultant control systems in terms of

stability and performance. We will employ two types of approaches, one of which is
based on μ analysis, i.e., the scaled H∞ -norm approach, and the other one utilizes
the state-dependent coefficient (SDC) form with the Lyapunov theory.
First, we present μ-based robustness analysis. As mentioned in the previous H∞
controller synthesis section, conventional μupper bounds with frequency-dependent
scalings D(iω)’s can accommodate only time-invariant uncertainties properly. For
time-varying uncertainties as in our case, constant scalings over all the frequencies
5
10
4
10
3
10
K
Gain
2b1red11
K
1b1red11 K
2 21
10 K
11
1
10
0
10 −1 0 1 2 3
10 10 10 10 10
Frequency (rad/s)
5
10
4
10
3
10
K
Gain
2b1red22
K
1b1red22 K
2 21
10 K
11
1
10
0
10 −1 0 1 2 3
10 10 10 10 10
Frequency (rad/s)
Fig. 4.18 Frequency responses of the H∞ controllers; solid line K 11 , dashed line K 21 , dashed-
dotted line K 1b1r ed11 (top) and K 1b1r ed22 (bottom), dotted line K 2b1r ed11 (top) and K 2b1r ed22
(bottom)
4
10
2
Ta
10
0 T
10
S
Gain
−2
10
Sp
−4
10
−6
10 −1 0 1 2 3
10 10 10 10 10
Frequency (rad/s)
4
10
2
T
a
10
0 T
10
S
Gain
−2
10
Sp
−4
10
−6
10 −1 0 1 2 3
10 10 10 10 10
Frequency (rad/s)
Fig. 4.19 Frequency responses of the resultant sensitivity functions with the respective H∞ con-
trollers (single-frequency case); solid line K 11 , dashed line K 21 , dashed-dotted line K 1b1r ed11 (top)
and K 1b1r ed22 (bottom), dotted line K 2b1r ed11 (top) and K 2b1r ed22 (bottom)
4
10
2
Ta
10
0 T
10
S
Gain
−2
10
Sp
−4
10
−6
10 −1 0 1 2 3
10 10 10 10 10
Frequency (rad/s)
4
10
2
T
a
10
0 T
10
S
Gain
−2
10
Sp
−4
10
−6
10 −1 0 1 2 3
10 10 10 10 10
Frequency (rad/s)
trollers (double-frequency case); solid line K 11 , dashed line K 21 , dashed-dotted line K 1b1r ed11 (top)
and K 1b1r ed22 (bottom), dotted line K 2b1r ed11 (top) and K 2b1r ed22 (bottom)
4
10
2
Ta
10
0 T
10
S
Gain
−2
10
Sp
−4
10
−6
10 −1 0 1 2 3
10 10 10 10 10
Frequency (rad/s)
4
10
2
T
a
10
0 T
10
S
Gain
−2
10
Sp
−4
10
−6
10 −1 0 1 2 3
10 10 10 10 10
Frequency (rad/s)
trollers (Bretschneidr case); solid line K 11 , dashed line K 21 , dashed-dotted line K 1b1r ed11 (top) and
K 1b1r ed22 (bottom), dotted line K 2b1r ed11 (top) and K 2b1r ed22 (bottom)
must be used for stringent analyses. Let G ev be the closed-loop system to be evaluated
which contains the plant, controller, maybe weighting functions, and Δev be the
time-varying normalized uncertainty with the compatible dimension with G ev . Then,
Fu (G ev , Δev ) is stable if and only if there exist Dl , Dr ∈ R which satisfy
Dl G ev Dr ∞ < 1 (4.90)
where Dl and Dr are constant nonsingular matrices with Dr Δev = Δev Dl−1 ( · ∞
denotes the H∞ -norm)(see, e.g., [68, 86, 93] for more detail). Then, let Dad denote
the admissible set of (Dl , Dr ) and define
μc := inf Dl G ev Dr ∞ . (4.91)
(Dl ,Dr )∈Dad
Therefore, robustness analysis must be done based on the criteria μc to accommodate

time-varying structured uncertainties. However, in terms of “control design,” it is
generally difficult to solve a problem via constant scalings-based D–K iteration [3,
79, 109] rather than conventional D–K iteration since the optimization problem
of minimizing μc is non-convex for output feedback H∞ control cases [109] and
constant scalings are strongly restrictive. Hence, in this work, we have employed
conventional D–K iteration for control design, but will deploy μc for robustness
analyses.
Regarding the obtained 12 controllers K ij , three types of robustness analyses
based on the respective generalized plants P , Paug1b , and Paug2b (see the Sect. 4.4.4
again) are conducted. P can accommodate only robust stability analysis without
any weighting functions for performances; Paug1b additionally contains Ws and Wu ,
and hence accommodates robust performance; and further Paug2b copes with all the
points, i.e., robust performance with Ws , Wu , and Wn . In the respective analyses,
G ev with Δev are set as Fl (P , K ij ) with Δ I ∈ I in (4.55), Fl (Paug1b , K ij ) with
Δ ∈ 1 in (4.82), and Fl (Paug2b , K ij ) with Δ ∈ 2 in (4.83), and μc ’s are obtained
by a numerical optimization method based on a quadratic programming algorithm
with the initial Dl and Dr of the identity matrices. The respective admissible sets are
⎧ ⎫
⎨ (Dl , Dr ) : Dl = blockdiag[D1 , D2 ], ⎬
D := Dr = blockdiag[D1−1 , D2−1 ], (4.92)
⎩ ⎭
Di (det Di
= 0) ∈ R2×2
for P , ⎧ ⎫
D1 := Dr = blockdiag[D1−1 , D2−1 , d1−1 I2 ], (4.93)
⎩ ⎭
Di (det Di
= 0) ∈ R2×2 , d1 (
= 0) ∈ R
Table 4.7 Results of μc with P for robust stability analyses against Δ I

j K1 j K 1bj K 21 K 2bj
1 0.5923 0.3693 0.5169 0.4780
0.3693a 0.4779a
2 0.6500 0.4645 0.5788 0.5693
0.4938a 0.5732a
3 0.6490 0.4640 0.5820 0.5315
0.4734a 0.5420a
1 single-frequency, 2 double-frequency, 3 Bretschneider.
a Denotes the one of the reduced-order controller
for Paug1b , and

⎧ ⎫
D2 := Dr = blockdiag[D1−1 , D2−1 , d1−1 I4 ], (4.94)
⎩ ⎭
Di (det Di
= 0) ∈ R2×2 , d1 (
= 0) ∈ R
for Paug2b .
Table 4.7 shows the results of μc ’s with P for robust stability analyses against
normalized time-varying structured model uncertainty Δ I in (4.55), where not only
the reduced-order controllers labeled with red but also the original controllers are
considered. First, it should be emphasized that all the μ’s are less than one, which
implies that all the controllers can ensure at least robust stability against Δ I . Then,
we compare μc ’s of K 1 j ’s and K 1bj ’s ( j = 1, 2, 3) and notice that μc ’s of K 1bj ’s are
less than those of K 1 j ’s. This fact reflects explicit consideration of model uncertain-
ties via extended matrix polytopes, and shows that the proposed model uncertainty
representation method works well. Further, the cases of K 2 j ’s and K 2bj ’s are similar,
but the differences of μc are smaller than those in the cases of K 1 j ’s and K 1bj ’s,
which may be due to some conflict between the model uncertainties and the weight-
ing function Wn . Next, consider the difference between μc ’s of K 1 j ’s and K 2 j ’s.
μc ’s of K 2 j ’s are smaller than those of K 1 j ’s, which is because Wn influences Ta
as well as S and Ta is associated with additive model uncertainties. Moreover, let
us compare K ibj ’s with the corresponding reduced-order controllers K ibjred ’s. In the
case of single-frequency oscillation, their μc ’s are almost the same, however μc ’s of
the reduced-order controllers are slightly larger than those of the original controllers.
Which shows that the model reduction process has decreased the robustness of the
controllers very slightly, but almost conserved it.
Table 4.8 presents the results of μc ’s with Paug1b without Wn for robust perfor-
mance analyses against Δ ∈ 1 . In the robust performance analyses, since it was
hard to obtain μc ’s of K ibj ’s, only those of K ibjred ’s are displayed. Not as in the case
of robust stability analyses, all the μc ’s are greater than one. However, which is not
a problem, because we have already confirmed that all the controllers can provide
robust stability. Further, those μc ’s are close to one, thus their robust performance
can be expected. With respect to μc ’s relationship among the controllers, the same
arguments can be applied as the above robust stability analyses.
Table 4.8 Results of μc with Paug1b without Wn for robust performance analyses against Δ ∈ 1
j K1 j K 1bjred K 21 K 2bjred
1 1.3431 1.0094 1.2373 1.0709
2 1.5621 1.4173 1.4951 1.4333
3 1.5502 1.3088 1.4884 1.3866
1 single-frequency, 2 double-frequency, 3 Bretschneider
Table 4.9 Results of μc with Paug2b with Wn for robust performance analyses against Δ ∈ 2
j K1 j K 1bjred K 21 K 2bjred
1 3.3209 3.4501 3.1337 2.7103
2 4.0071 4.3391 3.8031 3.9539
3 4.0114 4.3496 3.8881 3.7440
1 single-frequency, 2 double-frequency, 3 Bretschneider
Finally, the results of μc ’s with Paug2b with Wn for robust performance analyses
against Δ ∈ 2 are displayed in Table 4.9. Due to the restrictive Wn , all the μc ’s are
greater than those in the above cases. The μc ’s relationship among the controllers
reveals different features from those in the above cases. One feature is that μc ’s
of K 1 j ’s and K 1bjred ’s are greater than those of K 2 j ’s and K 2bjred ’s, which is due
to lacking of consideration of Wn . The other feature is that μc ’s of K 1bjred ’s are
greater than those of K 1 j ’s, which is not the case for the above cases, and implies
some conflict between Wn and the model uncertainties. On the other hand, μc ’s of
K 2bjred ’s are less than those of K 2 j ’s with one exception of the double-frequency
cases. Moreover, K 2bjred ’s can be expected to have good robust performance in such a
situation where model uncertainties and sensor noises need to be taken into account.
From all the above arguments, as long as the μ-based robust analysis is concerned,
we have concluded as follows:
• The proposed extended matrix-polytope-based model uncertainty representation
is feasible for both control design and robustness analyses;
• As a H∞ synthesis method with consideration of structured uncertainties, the
conventional D–K iteration is a useful and systematic way even if the uncertainties
are real and/or time-varying ones;
• Further, in such a case, the combination of the conventional D–K iteration and
μc -based robustness analyses is an effective approach;
• Wn and model uncertainties might cause some conflict, thus the balance of the
gains of Wn and Ws will be important in a control design procedure.
Now we proceed to state-dependent coefficient (SDC) form-based robust stability
analyses. The control system is constructed based on the virtual linear system in
(4.62), its stability is thus guaranteed for the linear system. However, once the payload
varies, the system is no longer linear. In the above μ-based analyses, the system
is considered the linear system with time-varying uncertainties. Here, we address
another viewpoint where the system is regarded as nonlinear focusing on its semi-
global internal stability, and apply a method that is motivated by [13].
As in [13], we utilize a SDC form as ẋ = A(x)x. When a state-space realization

of the mth-order H∞ controller is represented by
ξ̇ = Ac ξ + Bc q (4.95)
u = Cc ξ , (4.96)
the SDC form of the closed-loop control system with the state x T = [q, q̇, ξ ]T can
be expressed as follows:
⎡ ⎤ ⎡ ⎤⎡ ⎤
q̇ 0 I2 0 q
⎣ q̈ ⎦ = ⎣ A21 A22 A23 ⎦ ⎣ q̇ ⎦ (4.97)
ξ̇ Bc 0 Ac ξ
A21 = M(q)−1 (−F1 Mn (q) − ΔG(q)) (4.98)
A22 = M(q)−1 (−F2 Mn (q) − ΔC(q, q̇)) (4.99)
A23 = M(q)−1 Mn (q)Cc (4.100)
where Mn is the inertia matrix with the nominal parameters, [F1 , F2 ] are linear
gains in (4.62), ΔG(q)q and ΔC(q, q̇)q̇ denote the gravitational term and Coriolis–
centripetal term perturbations respectively.
Using this SDC form, we first investigated the eigenvalues λi ’s of A(x) over a
state-space region Ω. Taking into account of simulation and experimental cases for
demonstration, we designate Ω as
q × q̇ × ξ = [−π/2, π/2] × [−π, π ] × [−5, 5] × [−12, 12] × Rm . (4.101)
Then, we calculated the maximum of the real part of the eigenvalues
max Re(λi (A(x))) (4.102)

x∈Ω,i
for a given payload width, the results of which are shown by the top graph in Fig. 4.22.
In Fig. 4.22, each graph of oscillation case compares the respective controllers. As
seen from the graph, for every payload width, the maximum Re(λi (A(x))) is nega-
tive, which indicates that the control system is possibly internally stable over Ω for the
payload variations. Further, interestingly enough, K 1 j and K 2 j via γ -iteration exhibit
almost the same profiles which vary depending on the payload widths, whereas
Re(λi (A(x)))’s with K 1bjred and K 2bjred via D–K iteration are greater than those
with K 1 j and K 2 j , and invariant with the payload width, that is, some robustness.
Strictly speaking, this condition is not sufficient for its robust stability. Cloutier
et al. [13] presented a sufficient condition for the internal stability only for the spe-
cial SDC form with a symmetric A(x), which is however not the general case. Hence,
here we construct a new machinery that is applicable to more general SDC forms
including ours, which is described by the following theorem.
max Re(λi)
−2
−4
0 1 2 3 4 5 6 7 8 9 10
0
max Re(λi)
−2
−4
0 1 2 3 4 5 6 7 8 9 10
0
max Re(λi)
−2
−4
0 1 2 3 4 5 6 7 8 9 10
width (mm)
Fig. 4.22 Results of robust stability analyses in terms of maximum Re(λi (A(x))) over Ω in (4.101)
versus payload width; Top single-frequency oscillation, Middle double-frequency oscillation,
Bottom Bretschneider oscillation; solid line K 1 j , dashed line K 2 j , dashed-dotted line K 1bjred ,
dotted line K 2bjred
Theorem 4.5 For a given system represented by ẋ = A(x)x, suppose that A(0) is
diagonalizable, then there exists a constant nonsingular matrix U that diagonalizes
A(0) by U A(0)U −1 . If U is a complex matrix, then U is replaced by Re(U )+ I m(U ).
For a given state-space region Λ, if

max λi (U −1 )T A(x)T U T + U A(x)U −1 < 0, (4.103)
x∈Λ,i
then the system is asymptotically stable w.r.t x = 0 over Λ, thus Λ is contained in

the entire region of attraction.
Proof Let V (x) = x T U T U x be a Lyapunov function candidate. Since U is nonsin-

gular, V (x) is positive definite. Differentiating V (x) yields
V̇ (x) = ẋ T U T U x
= x T A(x)T U T U x + x T U T U A(x)x

= x T U T (U −1 )T A(x)T U T + U A(x)U −1 U x. (4.104)
q bound (rad)
2
2
0
0 1 2 3 4 5 6 7 8 9 10
4
q bound (rad)
2
2
0
0 1 2 3 4 5 6 7 8 9 10
4
q bound (rad)
2
2
0
0 1 2 3 4 5 6 7 8 9 10
width (mm)
Fig. 4.23 Results of robust stability analyses based on Theorem 4.5; q̄2 of the region of attrac-
tion versus payload width; Top single-frequency oscillation, Middle double-frequency oscillation,
Bottom Bretschneider oscillation; solid line K 1 j , dashed line K 2 j , dashed-dotted line K 1bjred , dotted
line K 2bjred
Then by (4.103) and U ’s nonsingularity V̇ (x) is negative definite over Λ. Hence the
proof is complete.
Based on the theorem, we calculated each Λ for each controller, that is, a subset
of the region of attraction that satisfies the condition (4.103) for a given payload
variation. The resulting Λ can be represented as
q × q̇ × ξ = [−π/10, π/10] × [−q̄2 , q̄2 ] × [−5, 5] × [−12, 12] × Rm , (4.105)
where q̄2 is the bound for q2 and varies with the payload variations, which is displayed
in Fig. 4.23. In Fig. 4.23, each graph is associated with each oscillation case, and
compares the respective controllers. As in the case of maximum Re(λi (A(x))), the
profiles with K 1 j and K 2 j via γ -iteration are almost the same, and quite similar
regardless of the oscillation cases. On the other hand, those with the controllers via
D–K iteration vary depending on the oscillation cases, however, reveal apparently
larger bounds than those with K 1 j ’s and K 2 j ’s, which show the stronger robustness.
This robust stability analysis method based on Theorem 4.5 is more stringent than
the above μ-based one, because any approximation on the evaluated system is not
used. This method is essentially based on the concept of “quadratic stability,” and
is very easy to implement and useful. However, the result provides only a sufficient
condition and hence can be conservative. In fact, as shown in Fig. 4.23, q̄2 ’s were
calculated to be zero in the cases of K 1b1r ed with the width of 0 mm and K 2b2r ed
with the width of 8 mm. In the sense of continuity, those results might not reflect the
actual properties and similar regions of attraction with those of near conditions. This
sort of singularity problem needs to be solved or compensated by other approaches.
Consequently, the SDC-based analyses have enhanced the conclusions of the
above μ-based robustness analyses from the different viewpoint. In particular, the
stringent stability analyses have proven that the proposed extended matrix-polytope-
based model uncertainty representation is useful for control design and the resulting
controller will have promising robustness against the model uncertainties.
4.5 Conclusions
In this section, we have presented the H∞ -control-based approach to the control

problem of OBMs. For this approach, we have developed the machinery, extended
matrix polytope, to efficiently and effectively model parametric uncertainties due to
the payload variation. The control method consists of nonlinear state-feedback con-
trol to reduce the nonlinearity of the system and the linear H∞ control scheme. In the
control design process, how to model uncertainties and what kinds of weighting func-
tions to be used for H∞ control design are important. Therefore, we have designed
four types of controllers with differences in uncertainty model and/or weighting func-
tions, and compared their properties by examining their frequency responses, poles,
and zeros and by performing robustness analysis with two types of tests. One of
the robustness analysis is the constant-scaling μ-analysis to accommodate the time-
varying uncertainties and the other one is based on the SDC form and the Lyapunov
theory. From all the results and arguments, we have concluded as:
• All the H∞ controllers have the corresponding frequency modes of the base oscilla-
tion as their poles, thus according to the internal model principle they are expected
to effectively reduce the disturbance due to the base oscillation;
• The weighting function Wn for sensor noise plays an important role in not only
noise reduction but also robustness enhancement;
• The extended matrix polytope can efficiently and effectively represent parametric
uncertainties due to the payload variation, and can provide stronger robustness to
the resultant controllers.
Again it should be noted that those results are not exclusive for the OBM control
problems but extensively applicable to wider class of mechanical system control
problems.
Chapter 5
Simulations and Experiments for the
H∞-Control-Based Approach
Abstract This chapter presents control performance evaluations of the H∞ -control-

based approach by simulations and hardware experiments separately from Chap. 4.
We have performed demonstrations of control performance considering the three
types of control problems and three patterns of base oscillations mentioned in Chap. 4,
and the four types of H∞ -controllers presented in Chap. 4. For those demonstration
cases, we investigate the respective control performance, with respect to nominal-
case performance, robust control performance against the payload variations, influ-
ence of sensor error, and moreover comparison with the conventional PID control.
Those results show that the H∞ -control-based approach and the extended matrix
polytope are practically useful and effective from the viewpoint of control perfor-
mance.
5.1 Introduction
This chapter presents simulation and experimental evaluations on the H∞ -control-

based approach presented in the previous chapter in terms of the control design
method, robustness analyses of the obtained control systems. As demonstration cases,
the three types of control problems, i.e., attitude control in local coordinates, attitude
control in global coordinates, position control in global coordinates, and the three
types of base oscillation, i.e., the single-frequency oscillation, the double-frequency
one, the Bretschneider one, are considered comparing the four types of H∞ con-
trollers.
This chapter is organized as follows. First, in Sect. 5.2 we present nominal-case
control performances, where examination of sensor error influence and effectiveness
of the weighting function Wn , and comparison with the conventional PID control are
included. Then, Sect. 5.3 demonstrates robust control performances. The last section
gives some conclusions.

DOI 10.1007/978-3-319-21780-2_5
68 5 Simulations and Experiments for the H∞ -Control-Based Approach
5.2 Simulations and Experiments (Nominal-Case

Performance)
First, we focus on nominal-case performances of the proposed approach comparing

the four types of controller K 1 j ’s, K 1bjred ’s, K 2 j ’s, and K 2bjred mentioned in the
previous chapter. Further, we will compare those controllers with a PID controller
to evaluate our approach fairly. We conducted simulations with respect to all the
combinations of three types of oscillation (single-frequency, double-frequency, and
Bretschneider oscillations), three types of control problems (attitude control in local
coordinates, attitude control in global coordinates, and position control in global
coordinates), and the four types of controller. Further, choosing some cases among
them, we investigated influence of sensor errors, comparison with the conventional
PID control, and carried out experiments using the experimental apparatus introduced
in Chap. 3. In all the demonstration cases presented in this section, the nominal case,
i.e., with the payload width of 5 mm is considered
5.2.1 Base Oscillation
Figure 5.1 shows the profiles of three types of base oscillation employed for demon-
strations, where the top graph depicts the single-frequency oscillation, the second
does the double-frequency one, and the bottom does the Bretschneider oscillation. In
10
qb (deg)
−10
0 2 4 6 8 10 12 14 16 18 20
10
q (deg)
0
b
−10
0 2 4 6 8 10 12 14 16 18 20
10
qb (deg)
−10
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.1 Base motion for simulations and experiments; top the single-frequency oscillation, middle
the double-frequency oscillation, bottom the Bretschneider oscillation
5.2 Simulations and Experiments (Nominal-Case Performance) 69
attitude control cases, the base starts oscillation at 4 s, while in position control case,
it starts at 10 s. In every case, the whole period of demonstration is 20 s. It should
be noted that the power of oscillation per frequency of the single-frequency case is
the largest one being compared with the others, which will be reflected by control
simulation results.
5.2.2 Global Coordinates
Here we define some variables for demonstrations. r1 and r2 denote the reference
signals to the respective manipulator joints, while rg1 and rg2 are the reference ones
in the global coordinates: rg1 = qb +r1 , rg2 = qb + q1 +r2 . Similarly, qg1 = qb + q1
and qg2 = qb + q1 + q2 are the angles of the links in the global coordinates.
5.2.3 PID Control
For comparison with the proposed control approach, the conventional PID control is
also demonstrated. The adopted PID controllers to the tracking error e are as in the
following:
t
τ1 = FP1 e1 + FD1 ė1 + FI 1 e1 dτ (5.1)
0
t
τ2 = FP2 e2 + FD2 ė2 + FI 2 e2 dτ (5.2)
0
where τ = [τ1 , τ2 ]T is the control torque vector, FPi , FDi , and FI i are the propor-
tional, derivative, and integral control gains for Link i respectively. As mentioned
below, those gains are determined by using the optimization scheme with a criterion
of root mean square control error by simulation similarly to the H∞ control case.
The optimization scheme for PID control is based on a quadratic programming algo-
rithm over [0.1, 10], [0.1, 10], [0.01, 1] for FI 1 , FP1 , FD1 and [0.01, 1], [0.01, 1],
[0.001, 0.1] for FI 2 , FP2 , FD2 respectively.
5.2.4 Attitude Control
For attitude control, we demonstrate two types of problems, i.e., in local coordinates
and in global coordinates respectively. The initial conditions are q1 (0) = q2 (0) =
qb (0) = 0 deg, q̇1 (0) = q̇2 (0) = q̇b (0) = 0 deg/s. The whole duration is 20 s. As
displayed in Fig. 5.1, the base starts oscillation at 4 s with the respective oscillations.
In the local coordinate demonstration, the references r1 and r2 are step function
such that r1 = 0 → 30 deg and r2 = 0 → 60 deg at 10 s. In the global coordinate
one, the references rg1 and rg2 are step functions such that rg1 = 0 → 30 deg and
rg2 = 0 → 90 deg at 10 s, that is, r1 = −qb → −qb + 30 and r2 = −q1 − qb →
−q1 − qb + 90 deg. To avoid overshoot, we applied a first-order low-pass filter with
the time constant 0.8 to those step signals.
As mentioned above, to optimize the linear state-feedback gains F = [F1 , F2 ],
the 401 time series angle tracking error vectors of simulation

e (16.00), e1 (16.01), . . . , e1 (20.00)
es = 1 (5.3)
e2 (16.00), e2 (16.01), . . . , e2 (20.00)
are adopted as the steady errors and the ||es || F is calculated as the criterion.
√ we defined the corresponding root mean square error (RMSE) as ē =
Further,
||es || F / 802, which aims at making us understand control performances intuitively.
By repeating simulations of the case of attitude control in global coordinates with
double-frequency oscillation under the above conditions, F for the H∞ controller and
PID gains the PID controller were optimized. To determine F, the H∞ controller K 22
was employed, and then the obtained F is commonly used for all the H∞ controllers.
As a result of the optimization,
• F1 = 66 and F2 = 24 were obtained;
• the optimal PID controller gains were as FI 1 = 2.9, FP1 = 2.3, FD1 = 0.05,
FI 2 = 0.19, FP2 = 0.32, FD2 = 0.01.
5.2.5 Position Control
Next, we demonstrate position control with respect to the center of the payload in
the global coordinates. With a contrast to attitude control, position control requires
an online solution to the inverse kinematics problem. Let (r x , r y ) denote the refer-
ence position of the center of the payload in the global coordinates. Then, the joint
references can be obtained as follows:

−1 L y
r x = L x + L y cos qb + tan
2 2 − rx
Lx

−1 L y
r y = r y − L x + L y sin qb + tan
2 2
Lx
r x2 + r y2 − l12 − l22
ψ= (5.4)
2l1l2

−1 1 − ψ2
r2 = tan
ψ

−1 r x −1 l2 sin r2
r1 = tan − tan − qb . (5.5)
r y l1 + l2 cos r2
The initial conditions are q1 (0) = q2 (0) = qb (0) = 0 deg, q̇1 (0) = q̇2 (0) =
q̇b (0) = 0 deg/s, which correspond to (r x (0), r y (0)) = (0.10, 0.20) (m). The whole
duration is 20 s. (r x , r y ) are two step reference signals such that (0.10, 0.20) →
(0.0826, 0.1988) at 4 s, which position corresponds to q1 (0) = qb (0) = 0 and
q2 (0) = 5 deg, and then (0.0826, 0.1988) → (0.06, 0.13) at 6 s, and the base starts
oscillation at 10 s, which are intended to avoid multiple solutions of the inverse
kinematics. To avoid overshoot, we applied the same low-pass filter as in the attitude
control case.
For controller parameter optimization, the 401 time series tracking “distance”
errors as es = [e(16.00), e(16.01), . . . , e(20.00)] are adopted by conducting simu-
lations with the double-frequency oscillation.
√ The corresponding mean of the steady-
state error was defined as ē = ||es || F / 401.
5.2.6 Simulation Results
Table 5.1 shows the results of RMSE ē in the respective control simulations. First, let
us focus on the differences among the base oscillation cases. As pointed out before,
ē’s in the case of single-frequency are larger those in the other cases, which is due
to the oscillation power per frequency. Then, we compare the types of problems. ē’s
of attitude control in global coordinates are larger than those of the local coordinate
Table 5.1 Results of RMSE ē in the respective control simulations

K 11 K 1b1red K 21 K 2b1red
Attitude-l (deg) 0.1081 0.1488 0.1047 0.1250
Attitude-g (deg) 0.2396 0.3463 0.2152 0.3309
Position (mm) 0.4080 0.8215 0.3972 0.6125
Attitude-l (deg) 0.1033 0.1236 0.1072 0.1047
Attitude-g (deg) 0.1723 0.2097 0.1361 0.1501
Position (mm) 0.2285 0.3443 0.2276 0.2521
Attitude-l (deg) 0.1059 0.1116 0.1085 0.1002
Attitude-g (deg) 0.1916 0.1955 0.1486 0.1424
Position (mm) 0.2393 0.4508 0.2358 0.3213
Attitude-l attitude control in local coordinates, attitude-g attitude control in global coordinates,
position position control
case. In the global coordinate problems, not only disturbance suppression but also
tracking control to oscillatory trajectory is required, which makes the problem more
difficult than the local coordinate ones.
Now we compare the types of controllers. Regarding the point of explicit model
uncertainty consideration, that is, comparing K ij and K ibjred , K ij performs better than
K ibjred . This result is natural since the demonstration is based on the nominal case.
Next let us investigate effect of the weighting function Wn for sensor noise. As those
results suggests, Wn is very effective in those demonstrations, which is because the
joint angle sensor resolution of 0.18 deg is rather coarse. This point will be later
discussed in more detail. In almost all the cases, K 2j ’s reveal the best performances,
while K 1bjred ’s do the worst performances. K 1 j ’s and K 2bjred ’s are in the middle class.
However, taking into account of the sensor resolution of 0.18 deg, those controllers
exhibit successful performances, which has proven that our proposed approach is
considerably effective and useful.
Next, in order to investigate time series data, we present the simulation results of
the proposed H∞ controllers in the cases of the single-frequency oscillation, which
can demonstrate the essential features of the four types of controllers and the three
types of control problems, because the oscillation power per frequency is the largest
as pointed out above. Figures 5.2, 5.3, 5.4, 5.5, 5.6, 5.7, 5.8, 5.9, 5.10, 5.11, 5.12, and
5.13 show the time series data of the control simulation results, respectively, in the
100
Link 1
qi (deg)
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.2 Simulation results of attitude control in local coordinates with the controller K 11 for the
single-frequency oscillation; top the attitudes of the links in the local coordinates, second from the
top the tracking errors on a normal scale, third from the top the tracking error of Link 1 on a fine
scale, bottom the tracking error of Link 2 on a fine scale
100
Link 1
q (deg)
50 Link 2
0
i
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
0.5
error (deg)
Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
0.5
error (deg)
Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.3 Simulation results of attitude control in local coordinates with the controller K 1b1red for
the single-frequency oscillation; top the attitudes of the links in the local coordinates, second from
the top the tracking errors on a normal scale, third from the top the tracking error of Link 1 on a
fine scale, bottom the tracking error of Link 2 on a fine scale
100
Link 1
qi (deg)
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
single-frequency oscillation; top the attitudes of the links in the local coordinates, second from the
100
Link 1
q (deg)
50 Link 2
0
i
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.5 Simulation results of attitude control in local coordinates with the controller K 2b1red for
the single-frequency oscillation; top the attitudes of the links in the local coordinates, second from
100
Link 1
qgi (deg)
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.6 Simulation results of attitude control in global coordinates with the controller K 11 for the
single-frequency oscillation; top the attitudes of the links in the global coordinates, second from
100
Link 1
qgi (deg)
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.7 Simulation results of attitude control in global coordinates with the controller K 1b1red for
the single-frequency oscillation; top the attitudes of the links in the global coordinates, second from
100
Link 1
qgi (deg)
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
single-frequency oscillation; top the attitudes of the links in the global coordinates, second from
100
Link 1
qgi (deg)
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.9 Simulation results of attitude control in global coordinates with the controller K 2b1red for
the single-frequency oscillation; top the attitudes of the links in the global coordinates, second from
150
Link 1
qgi (deg)
100
Link 2
50
0
0 2 4 6 8 10 12 14 16 18 20
0.1
error (m)
0.05
0
0 −3 2 4 6 8 10 12 14 16 18 20
x 10
5
error (m)
0
0 2 4 6 8 10 12 14 16 18 20
torque (Nm)
0.1
Link 1
0 Link 2
−0.1
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.10 Position-control simulation results with the controller K 11 for the single-frequency oscil-
lation; top the attitudes of the links in the global coordinates, second from the top the distance errors
from the final goal on a normal scale, third from the top the distance errors from the final goal on a
fine scale, bottom the control torques
150
Link 1
q (deg)
100
Link 2
50
gi 0
0 2 4 6 8 10 12 14 16 18 20
0.1
error (m)
0.05
0
0 2 4 6 8 10 12 14 16 18 20
−3
x 10
5
error (m)
0
0 2 4 6 8 10 12 14 16 18 20
torque (Nm)
0.1
Link 1
0 Link 2
−0.1
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.11 Position control simulation results with the controller K 1b1red for the single-frequency
oscillation; top the attitudes of the links in the global coordinates, second from the top the distance
errors from the final goal on a normal scale, third from the top the distance errors from the final
goal on a fine scale, bottom the control torques
150
Link 1
qgi (deg)
100
Link 2
50
0
0 2 4 6 8 10 12 14 16 18 20
0.1
error (m)
0.05
0
0 −3 2 4 6 8 10 12 14 16 18 20
x 10
5
error (m)
0
0 2 4 6 8 10 12 14 16 18 20
torque (Nm)
0.1
Link 1
0 Link 2
−0.1
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.12 Position control simulation results with the controller K 21 for the single-frequency oscil-
lation; top the attitudes of the links in the global coordinates, second from the top the distance errors
from the final goal on a normal scale, third from the top the distance errors from the final goal on a
fine scale, bottom the control torques
150
Link 1
q (deg)
100
Link 2
50
gi 0
0 2 4 6 8 10 12 14 16 18 20
0.1
error (m)
0.05
0
0 −3 2 4 6 8 10 12 14 16 18 20
x 10
5
error (m)
0
0 2 4 6 8 10 12 14 16 18 20
torque (Nm)
0.1
Link 1
0 Link 2
−0.1
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.13 Position control simulation results with the controller K 2b1red for the single-frequency
order of attitude control in local coordinates, attitude control in global coordinates,

and position control.
First, we consider the results of attitude control in local coordinates in Figs. 5.2,
5.3, 5.4, and 5.5. Each figure consists of four graphs. The top graph displays attitudes
of the two links in the local coordinates. The second graph shows tracking control
errors of the joints on a normal scale. Further, the third and fourth ones magnify
the respective control error on a fine scale. Viewing the top graph in each figure,
all the results similarly exhibit good control performances. But, seeing the tracking
error profiles, we see slight differences among the controllers. See Fig. 4.18 which
represents the frequency responses of the respective controllers K 11 –K 2b1red , and
recall that the gains of K 1b1red are the lowest and those of K 2b1red are the second
lowest. Due to the fact, K 1b1red requires slightly longer time to reach the steady state
and reveals slightly poorer steady-state performance than the other controllers. K 11
and K 21 exhibit almost the same best performances.
Then, we proceed to the results of attitude control in global coordinates in Figs. 5.6,
5.7, 5.8, and 5.9. The components of each figure are almost the same as those of
the figure of local coordinate problem, except that the top graph is based on the
global coordinates. It is seen from the figures that the essential features in terms
of differences among the four types of controllers are the same as in the above
arguments, and are even more manifest. This fact is also suggested by the data of ē’s
in Table 5.1.
Next, we present the results of position control in global coordinates in Figs. 5.10,
5.11, 5.12, and 5.13. In each figure, the top graph displays the attitude of the two links
in the global coordinates; the second represents the distance error from the final goal
position; the third one magnifies the distance error on a fine scale; and the bottom
graph does the control torques of the respective joints. Again, these results reveal the
same features in terms of differences among the controllers.
Finally, we see the simulation results of control torques. Figures 5.14 and 5.15
show the control torques of the joints 1 and 2, respectively, with K 11 . Note that
the profiles of control torques are essentially the same in the cases of the other
controllers. As in the order of control problems in the figures, control efforts are
most required in the attitude control in local coordinates, and the attitude control in
global coordinates is the second. Regarding the differences in the power of control
torques and the differences in the resulting control performances, it seems that in
the control problem where the less power of control torques are required, the gap
between the resultant control performances of K ij and K ibjred is the larger.
Consequently, by investigating all the simulation results carefully, we have pre-
sented the fundamental properties of the proposed H∞ controllers and their rela-
tionships in the three types of control problems. As long as the nominal situation is
concerned, all the proposed controllers work effectively, in particular the controller
K 2 j ’s perform best.
0.2
torque (Nm)
−0.2
0 2 4 6 8 10 12 14 16 18 20
0.2
torque (Nm)
−0.2
0 2 4 6 8 10 12 14 16 18 20
0.2
torque (Nm)
−0.2
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.14 Control torques of the joint 1 with K 11 for the respective problems in the case of single-
frequency oscillation; top attitude control in local coordinates, middle attitude control in global
coordinates, bottom position control in global coordinates
0.02
torque (Nm) 0
−0.02
0 2 4 6 8 10 12 14 16 18 20
0.02
torque (Nm)
−0.02
0 2 4 6 8 10 12 14 16 18 20
0.02
torque (Nm)
−0.02
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.15 Control torques of the joint 2 with K 11 for the respective problems in the case of single-
frequency oscillation; top attitude control in local coordinates, middle attitude control in global
coordinates, bottom position control in global coordinates
5.2.7 Influence of Sensor Error
We here investigate influence of sensor error to the resulting control performances,

in particular emphasizing the effectiveness of the weighting function Wn . To this
end, we employ the simulation results of attitude control in global coordinates in the
case of double-frequency oscillation. In order to focus on differences between with
Wn and without Wn , we compare the results of K 12 and K 22 .
Figures 5.16 and 5.17 compare the simulation results between with the low sensor
resolution 0.18 deg and a high sensor resolution 0.0018 deg, i.e., 100 times finer, in the
case of K 12 and K 22 respectively. In each figure, the upper two graphs show the results
with the low sensor resolution and the lower ones do those with the high resolution.
First, consider the case of the low resolution, where K 22 with consideration of Wn
in control design reveals better performances than K 12 without Wn , in particular
for Link 2. The resulting RMSE ē with K 22 is 0.1361 deg, and the one with K 12 is
0.1723 deg. Then, regrading the case of the high resolution, conversely K 12 performs
slightly better with ē = 0.0636 deg than K 22 with ē = 0.0675 deg, which might be
difficult to be seen in these graphs.
To make the argument more obvious, we present the power spectral density of the
steady-state control errors (time = 16.00–20.00) in Figs. 5.18 and 5.19. Figure 5.18
compare those situations for Link1, and Fig. 5.19 for Link 2. First, focus on the peaks
at very low frequencies, that is, 0.5 and 1 Hz, which have appeared in all the graphs
0.5
error (deg)
Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
0.5
error (deg)
Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
0.5
error (deg)
Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
0.5
error (deg)
Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.16 Comparison of sensor resolution in attitude control in global coordinates with K 12 in
the case of double-frequency oscillation; the upper two graphs: simulation results with the sen-
sor resolution of 0.18 deg, the lower two graphs: simulation results with the sensor resolution of
0.0018 deg
0.5
error (deg)
Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
0.5
error (deg)
Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
0.5
error (deg)
Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
0.5
error (deg)
Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.17 Comparison of sensor resolution in attitude control in global coordinates with K 22 in
the case of double-frequency oscillation; the upper two graphs: simulation results with the sen-
sor resolution of 0.18 deg, the lower two graphs: simulation results with the sensor resolution of
0.0018 deg
° °
K12(0.18 ) K22(0.18 )
0.01 0.01
0.008 0.008
0.006 0.006
PSD
PSD
0.004 0.004
0.002 0.002
0 0
0 10 20 30 40 50 0 10 20 30 40 50
°
K12(0.0018 ) K22(0.0018°)
0.01 0.01
0.008 0.008
0.006 0.006
PSD
PSD
0.004 0.004
0.002 0.002
0 0
0 10 20 30 40 50 0 10 20 30 40 50
frequency (Hz) frequency (Hz)
Fig. 5.18 Power spectral density (PSD) of the steady-state control errors (time = 16.00–20.00) of
the joint 1
K12(0.18°) K22(0.18°)
0.04 0.04
0.03 0.03
PSD
PSD
0.02 0.02
0.01 0.01
0 0
0 10 20 30 40 50 0 10 20 30 40 50
°
K (0.18 ) K (0.0018°)
12 22
0.04 0.04
0.03 0.03
PSD
PSD
0.02 0.02
0.01 0.01
0 0
0 10 20 30 40 50 0 10 20 30 40 50
frequency (Hz) frequency (Hz)
Fig. 5.19 Power spectral density (PSD) of the steady-state control errors (time = 16.00–20.00) of
the joint 2
and represents the residual errors due to the base oscillation. Interestingly enough, in
the case of the low resolution (see the upper graphs), although the RMSE ē with K 22 is
less than that with K 12 , its component due to the base oscillation with K 22 is greater
than that with K 12 . Further, it is seen that K 12 suffers from more high-frequency
noises due to sensor errors than K 22 . Then, in the case of the high resolution (see
the lower graphs), the components due to sensor errors have disappeared, and K 12
exhibits less power of control error than K 22 .
From the above argument, we see that the weighting function Wn for sensor noise
in H∞ control design plays an important role in such a case the sensor noise is a
problem, however, slightly sacrifices the tracking control performance of the control
system.
5.2.8 Experimental Results and Comparison with PID Control
Focusing on attitude control in global coordinates with the double-frequency oscilla-

tion, we compare the simulation results and experimental ones, and further compare
the proposed approach with the conventional PID control. For this demonstration,
the H∞ controller K 22 , which reveals the best performance in the nominal case, has
been chosen. And it should be noted that the PID gains have been optimized for this
control problem as described above.
First, we compare the simulation and experimental results with K 22 as shown
in Figs. 5.20 and 5.21 respectively. These figures have the same structure as the
corresponding ones in the case of single-frequency oscillation in the previous section.
Then, it is seen that not only the simulation results but also the experimental ones
exhibit successful control performances and both the profiles in the simulation and
experimental results are considerably consistent, although the experimental control
errors for Link 2 are slightly noisier than the simulation ones. The RMSE ē of
simulation is 0.1361 deg, while ē of experiment is 0.1842 deg. Additionally, these
results have proven that the obtained parameters in the dynamical model of the
experimental apparatus are accurate and the apparatus are appropriately developed.
Next, we present the simulation and experimental results with the PID controller.
Figures 5.22 and 5.23 display the simulation and experimental results with the PID
controller respectively. Concentrating on the steady-state control errors after the start
of step tracking control at 10 s in the simulation results in Fig. 5.22, the PID controller
performs successfully and better with ē = 0.1142 deg than K 22 with ē = 0.1361 deg.
However, unfavorable residual oscillations due to the base oscillation have appeared
on the control errors from 4 to 10 s for both the links. This difference has come from
the difference in the attitudes in the both cases, and suggests that the performance of
the PID controller might largely vary depending on the manipulator attitudes. Recall
that the PID gains have been tuned by taking only ē into account. Moreover, let us
see the experimental results in Fig. 5.23. Then, it is seen that Link 2 suffers from
tremendous large oscillation, which implies that the proportional and/or derivative
gains of the PID controller are too large under the experimental environments.
100
Link 1
qgi (deg)
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
double-frequency oscillation; top the attitudes of the links in the global coordinates, second from
100
Link 1
q (deg)
50 Link 2
gi
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
Fig. 5.21 Experimental results of attitude control in global coordinates with the controller K 22
for the double-frequency oscillation; top the attitudes of the links in the global coordinates, second
from the top the tracking errors on a normal scale, third from the top the tracking error of Link 1
on a fine scale, bottom the tracking error of Link 2 on a fine scale
100
Link 1
qgi (deg)
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.22 Simulation results of attitude control in global coordinates with the PID controller for the
double-frequency oscillation; top the attitudes of the links in the global coordinates, second from
100
Link 1
q (deg)
50 Link 2
gi
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
Fig. 5.23 Experimental results of attitude control in global coordinates with the PID controller
for the double-frequency oscillation; top the attitudes of the links in the global coordinates, second
from the top the tracking errors on a normal scale, third from the top the tracking error of Link 1
Furthermore, we demonstrate the comparison between K 22 and the PID controller

for the other control problems. Figures 5.24 and 5.25 depict the simulation results of
position control for the double-frequency oscillation with K 22 and the PID controller
respectively. In this control problem, both the controllers perform well. But, K 22 is
better with ē = 0.2276 mm than the PID controller with ē = 0.3238 mm, and the con-
trol torques of the PID controller for Link 2 are noisier than those of K 22 . Figures 5.26
and 5.27 then present the simulation results of attitude control in local coordinates
for the double-frequency oscillation with K 22 and the PID controller respectively.
In this case, the PID controller suffers from considerably poor performance with
ē = 0.3857 deg, while K 22 performs successfully with ē = 0.1072 deg.
Consequently, we have confirmed that the proposed H∞ controller is much supe-
rior to the PID controller with respect to “robustness.” Although PID control enjoys
the advantages of easy design and implementation, and reasonable performance with
appropriate gain tuning, as demonstrated above, the obtained single PID controller
cannot be flexible or robust to variations of situation and/or environment. On the other
hand. the proposed H∞ controller can perform successfully against such variations,
and thus has strong robustness. Table 5.2 summarizes ē’s in those demonstrations.
150
Link 1
qgi (deg)
100
Link 2
50
0
0 2 4 6 8 10 12 14 16 18 20
0.1
error (m)
0.05
0
0 −3 2 4 6 8 10 12 14 16 18 20
x 10
5
error (m)
0
0 2 4 6 8 10 12 14 16 18 20
torque (Nm)
0.1
Link 1
0 Link 2
−0.1
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.24 Position control simulation results with the controller K 22 for the double-frequency
150
Link 1
q (deg)
100
Link 2
50
gi 0
0 2 4 6 8 10 12 14 16 18 20
0.1
error (m)
0.05
0
0 −3 2 4 6 8 10 12 14 16 18 20
x 10
5
error (m)
0
0 2 4 6 8 10 12 14 16 18 20
torque (Nm)
0.1
Link 1
0 Link 2
−0.1
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.25 Position control simulation results with the PID controller for the double-frequency
100
Link 1
qi (deg)
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
double-frequency oscillation; top the attitudes of the links in the local coordinates, second from the
100
Link 1
q (deg)
50 Link 2
0
i
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.27 Simulation results of attitude control in local coordinates with the PID controller for the
double-frequency oscillation; top the attitudes of the links in the local coordinates, second from the
Table 5.2 Results of RMSE ē in demonstrations for comparison between K 22 and the PID control
K 22 PID
Attitude-l (deg) 0.1072 0.3857
Attitude-g (deg) 0.1361 0.1142
Attitude-g-ex (deg) 0.1842 0.2303
Position (mm) 0.2276 0.3238
Attitude-l attitude control in local coordinates, attitude-g attitude control in global coordinates
(-ex denotes the experimental results)
5.3 Simulations and Experiments (Robust Performance)
5.3.1 Robust Control Simulations and Experiments
To evaluate robustness of the proposed H∞ controllers, simulations for all the com-
binations of the four types of controllers, the three types of control problems, and
the three types of base oscillations were conducted. In addition to them, robust con-
trol experiments for selected cases among them were also conducted in comparison
with the PID controller. In order to implement physical parametric variations of the
manipulator, all the demonstrations deployed payload exchanges. As described in
Chap. 3, the payload of the experimental OBM is exchangeable with its width of
5.3 Simulations and Experiments (Robust Performance) 89
0, 1, 2, . . . , 10 mm (See Table 3.4 again to understand how those width variations

induce physical parametric variations).
Except the payload, the simulation and experimental methods are the same as those
of the nominal-case ones in Sect. 5.2. For robust control experiments, being based
on the attitude control in global coordinates for the double-frequency oscillation,
the H∞ controllers K 22 and K 2b2r ed , and the PID controller were employed. One
of the objectives of the experiments is to investigate effectiveness of the explicit
consideration of model uncertainties in control design, that is, to see how robust
K 2b2r ed can be in the experimental environment.
5.3.2 Results
The results of the robust control demonstrations are mainly represented and evaluated
by using RMSE ē’s. Figures 5.28, 5.29, and 5.30 show graphs of ē versus payload
width in the respective oscillation cases. Each figure consists of the graph of attitude
control in local coordinates (top), the one of attitude control in global coordinates
(middle), and the one of position control in global coordinates (bottom).
0.4
RMSE (deg)
0.2
0
0 1 2 3 4 5 6 7 8 9 10
0.4
RMSE (deg)
0.2
0
0 1 2 3 4 5 6 7 8 9 10
RMSE (mm)
0.8
0.6
0.4
0.2
0 1 2 3 4 5 6 7 8 9 10
payload width (mm)
Fig. 5.28 RMSE ē versus payload width of the simulation results in the case of single-frequency
oscillation; top attitude control in local coordinates, middle attitude control in global coordinates,
bottom position control in global coordinates; solid K 11 , dashed K 1b1r ed , dashed-dotted K 21 , dotted
K 2b1r ed
0.4
RMSE (deg) 0.2
0
0 1 2 3 4 5 6 7 8 9 10
0.4
RMSE (deg)
0.2
0
0 1 2 3 4 5 6 7 8 9 10
RMSE (mm)
0.8
0.6
0.4
0.2
0 1 2 3 4 5 6 7 8 9 10
payload width (mm)
Fig. 5.29 RMSE ē versus payload width of the simulation results in the case of double-frequency
bottom position control in global coordinates; solid K 12 , dashed K 1b2red , dashed-dotted K 22 , dotted
K 2b2r ed
0.4
RMSE (deg)
0.2
0
0 1 2 3 4 5 6 7 8 9 10
0.4
RMSE (deg)
0.2
0
0 1 2 3 4 5 6 7 8 9 10
RMSE (mm)
0.8
0.6
0.4
0.2
0 1 2 3 4 5 6 7 8 9 10
payload width (mm)
Fig. 5.30 RMSE ē versus payload width of the simulation results in the case of Bretschneider
bottom position control in global coordinates; solid K 13 , dashed K 1b3red , dashed-dotted K 23 , dotted
K 2b3r ed
Here we should emphasize that, in the nominal case, the essential features and
relationships of the four types of controllers do not vary with the oscillation cases,
however, which is not the case in the robust control demonstrations. First, let us
see Fig. 5.28 representing the single-frequency oscillation case, where ē’s of every
controller are invariant over the payload widths and thus the relationships among the
controllers are the same as those in the nominal case. Then, carefully investigating
Figs. 5.29 and 5.30, we see that the relationships of the controllers are different from
those in Fig. 5.28 and in the nominal case, in the attitude control cases and with the
smaller widths. In particular, in the graphs of attitude control in global coordinates
in both the cases of double-frequency and Bretschneider oscillations, K ibjred reveals
better performance than the corresponding K ij , which show effectiveness of the
explicit consideration of model uncertainties in control design by using extended
matrix polytopes. This effectiveness becomes the more apparent when the oscillation
pattern becomes the more complicated. K 2b2red and K 2b3red have strong robustness
and reveal the best worst-case performances in such situations.
On the other hand, one question has arisen to us. Considering the results of the
position control in global coordinates, the relationships of all the controllers are
completely invariant against such model variations in all the cases of base oscillations,
although the essential control schemes required are similar with those for the attitude
control in global coordinates. Then, where does such difference come from? At the
moment, we have not got a clear answer to the question. However, since a stark
contrast between the two control problems is the difference in the power of required
control torques as displayed in Figs. 5.14 and 5.15, we suppose this point lead to the
difference in the relationships of the controllers. If the force applied to the mass is
the larger, variation in the mass leads to the larger change in the motion, hence to
keep the motion in such a case, the more robustness is necessary for the controller.
Next, we address the experimental results. Figure 5.31 shows RMSE ē versus
payload width of the experimental results of attitude control in global coordinates
in the case of double-frequency oscillation similarly with those of simulations, but
includes the results of the PID controller. In the experiments, the payload widths of
0, 2, 5, 8, 10 mm were adopted. Each trial was repeated three times, and the mean
of ē’s was obtained as the final datum. As seen from the figure, K 22 and K 2b2red
reveal strong robustness even in the experimental situations, while the performance
of the PID controller dramatically deteriorates at width of 2 mm. As shown later, the
experimental OBM with the PID control under such a condition exhibited unfavorable
chattering phenomenon. Hence we avoided conducting the trial with the width of
0 mm for the PID controller.
Figures 5.32, 5.33, and 5.34 display the control error profiles of Link 2 on a fine
scale over the payload width variations with K 22 , K 2b2r ed , and the PID controller
respectively. Figures 5.32 and 5.33 show that both K 22 and K 2b2 can provide strongly
robust and successful performances over such payload variations. However, by care-
fully checking those figures, in Fig. 5.32 the profiles at widths of 8 and 10 mm are
slightly different from those at widths of 2 and 5 mm, whereas in Fig. 5.33 all the pro-
files are almost the same. Thus, K 2b2r ed is more robust than K 22 . On the other hand,
1.2
K
22
K
2b2red
1 PID
0.8
RMSE (deg)
0.6
0.4
0.2
0
0 1 2 3 4 5 6 7 8 9 10
payload width (mm)
Fig. 5.31 RMSE ē versus payload width of the experimental results of attitude control in global
coordinates in the case of double-frequency oscillation; solid K 22 , dashed K 2b21r ed , dotted PID
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.32 Experimental results of attitude control in global coordinates for Link 2 with K 22 over
the payload width variations; top 10 mm, second: 8 mm, third: 5 mm, bottom 2 mm
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.33 Experimental results of attitude control in global coordinates for Link 2 with K 2b2r ed
over the payload width variations; top 10 mm, second: 8 mm, third: 5 mm, bottom 2 mm
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 5.34 Experimental results of attitude control in global coordinates for Link 2 with the PID
controller over the payload width variations; top 10 mm, second: 8 mm, third: 5 mm, bottom 2 mm
Fig. 5.34 expresses poor robustness of the PID controller, in particular the bottom
graph at width of 2 mm exhibit unfavorable chattering phenomenon.
Consequently, all those results and arguments have proven that the proposed con-
trollers have strong robustness against payload variations and are also superior to
the PID controller in robustness. With respect to the explicit consideration of model
uncertainties in control design, its effectiveness has been confirmed, and its necessity
depends on the problems and situations. It should be noted that the demonstration of
the model uncertainty representation via the proposed extended matrix polytope in
this monograph is rather conservative for the simulation and experimental examples,
because the range of q2 is assumed to be [−π, π] in the model uncertainty represen-
tation but its range in the simulations and experiments is within [−π/2, π/2].
5.4 Conclusions
In this chapter, we have demonstrated control performances of the designed H∞

controllers by simulations and experiments. As demonstration cases, the three types
of control problems and the three types of base oscillations have been considered.
In the respective demonstration cases, we have compared the four types of H∞
controllers, i.e., with Wn or not, and with explicit consideration of model uncertainties
or not, in both the nominal case and robust control case. Additionally, we have
examined the influence of sensor error and compared the H∞ controllers with the
conventional PID controller. Consequently, the following conclusions have been
obtained:
• In all the demonstration cases, that is, regardless the types of problems and base
oscillations, all the H∞ controllers have revealed successful performance with
respect to both tracking control and disturbance suppression;
• the H∞ controllers are superior to the PID controller in terms of nominal control
performance and robustness. In particular, in the robustness, there is a stark contrast
between those controllers;
• In accordance with the results of analyses in the previous chapter, it has been
confirmed that the weighting function Wn is very effective to reduce the influence
of sensor error, and also plays an important role to enhance the robustness of the
control system;
• In comparison of the four types of H∞ controllers, the controller with Wn and with-
out explicit consideration, denoted by K 2 j ’s, have revealed the best performances
in most of the demonstration cases;
• With respect to the effectiveness of explicit consideration of model uncertainties
via extended matrix polytope, in robust control cases, the corresponding controllers
have exhibited strong robustness. The control performances of the controllers with-
out Wn are poorer than those of K 2 j ’s, however, by adding Wn the performances
can be greatly improved to be near to those of K 2 j ’s. Furthermore, in some robust
control cases, these controllers have the best performances. Thus, those results have
5.4 Conclusions 95
convinced us that the extended matrix polytope is effective not only in robustness
analysis but also in control design.
It should be noted that the oscillatory base problems obviously contains ordinary
problems, i.e., with non-oscillatory bases, as a special class. Therefore, the proposed
methodology with some tools such as an extended polytope is extensively applicable
to general control problems of mechanical systems.
Chapter 6
Motion Control Using a
Sliding-Mode-Control-Based Approach
Abstract In this chapter, we present the other highlight of the monograph, the
sliding-mode control (SMC)-based approach for the OBM robust control problems.
In our attempt to apply the SMC framework to such problems, we have developed
a novel nonlinear sliding surface, which is called the “rotating sliding surface with
variable-gain integral control (RSSI).” The advantageous feature of this method is
the control system can achieve successful tracking control and disturbance rejection
in not only steady state but also transient state with less control inputs. We present
the control design method, stability analyses on the RSSI and control performance
demonstrations by simulations in comparison with those of the H∞ -control-based
approach. The results show that, in the ideal situation in terms of sensor resolu-
tion, sampling period, control input limitation, the SMC-RSSI system is consider-
ably superior to the H∞ controllers, however once such ideal conditions have been
violated, its performance are dramatically deteriorated. Therefore, the SMC-based
approach is promising, but needs to be improved in the sense of practical implemen-
tation.
6.1 Introduction
In this chapter, we present the other proposed control design approach, that is, a slid-
ing mode control (SMC)-based one. Theoretically, SMC can be expected to achieve
good robustness and control performance without a priori knowledge on frequen-
cies of the base oscillation [5, 20, 39, 41, 48, 60, 73, 102, 103, 111], although there
exist some serious gaps between the theory and practical implementations. Thus, it is
worth applying it to the control problem of OBMs. Furthermore, we have proposed
a novel approach based on SMC by introducing a nonlinear sliding surface with
variable-gain integral control, which will be a powerful tool to provide advantageous
features with respect to transient-state control performance and control inputs when
compared with the conventional SMC.

DOI 10.1007/978-3-319-21780-2_6
98 6 Motion Control Using a Sliding-Mode-Control-Based Approach
This chapter is organized as follows. Section 6.2 presents a brief introduction to the
conventional SMC. In Sect. 6.3, we introduce our SMC-based approach and conduct
stability analysis of the nonlinear sliding surface using the illustrative model as in
the preceding chapters. Section 6.5 presents simulations to evaluate the proposed
control system by comparing with a conventional SMC system, and the proposed
H∞ control-base approach in Chaps. 4 and 5, and finally, some concluding remarks
are given.
6.2 Sliding-Mode Control
In this section, we briefly review the SMC framework. The SMC belongs to the wider
class of variable structure systems (VSS), and its origin started in 1950s in USSR.
Then, since Itkis [41] and Utkin [101] introduced the SMC theory to the world in the
late 1970s, the SMC has gained a lot of attentions and become to be popular. Now,
the SMC is well known to be one of the most powerful tools for robust control.
The SMC has unique features in the control law and control system structure, as
the name of VSS suggests. The control law is a nonlinear and discontinuous function
of the feedback variable, specifically a switching function. The system structure
consists of two different structures called “the reaching mode” and “the sliding
mode.” Correspondingly to the structure, the design procedure also consists of two
steps. First, the stable manifold in the state space of the system should be designed,
on which manifold the trajectory necessarily converges to the origin asymptotically.
Second, the control law to drive the system trajectory to the stable manifold and
thereafter constrain it to the manifold should be chosen. Once the trajectory has been
successfully confined to the manifold, the trajectory will automatically converge to
the origin, like sliding on the manifold, which is the origin of the name, and the stable
manifold is called the sliding surface.
In theory, the SMC is s considerably effective method for robust control, that
is, to overcome problems due to disturbances and model uncertainties. However,
there exist serious gaps between the theory and practical applications mainly due to
the discontinuous nature of switching low which requires infinite-velocity switching.
Hence, a direct implementation of the SMC in practice may often lead to unfavorable
chattering phenomena and even worse instability of the system. Therefore, a lot
of research efforts have been devoted to reduce or eliminate such gaps [25, 26],
while other researchers have been trying to improve the SMC itself by considering
a sliding surface, such as higher order [56], time-varying [11, 12], and nonlinear
sliding surfaces [5, 32]. Our work also belongs to the latter works.
6.3 Sliding-Mode Control via Rotating Sliding Surface … 99
6.3 Sliding-Mode Control via Rotating Sliding Surface

with Variable-Gain Integral Control
In this section, we address our proposed approach of sliding-mode control aiming at

application to systems subject to strong disturbances such as an OBM. The approach
is based on a nonlinear sliding surface proposed in [32] and additionally employs
variable-gain integral control which will enhance the robustness of the control system
against disturbances. The nonlinear sliding surface by [32] is originally motivated
by the works of [11, 12] which proposed a sliding surface with the time-varying
slope on the second-order phase plane, which is called a “rotating sliding surface.”
Reference [32] has arranged this idea by nonlinear tuning and inherited the term
“rotating.” Therefore, we also follow this stream and use the term “rotating sliding
surface,” when referring to the nonlinear sliding surface in our approach. Further, we
refer to the rotating sliding surface with variable-gain integral control as RSSI, SMC
via RSSI as SMC-RSSI, and conventional SMC via constant-gain integral control as
SMC-I, respectively, for abbreviation. Those abbreviations will be used for both the
respective control schemes and controllers using the schemes.
In this section, we present control system design of SMC-RSSI for the illustrative
model of OBM mentioned in Chaps. 2 and 3.
6.3.1 Control System Design of SMC-RSSI
Here, we explain about control system design of SMC-RSSI using the problem setting
in Chap. 2 and analyze a condition under which the convergence to the sliding surface
is ensured, that is, “a sliding condition.” Since the manipulator has two links, we
design two SMC-RSSI systems for Links 1 and 2, respectively.
The dynamical model of the manipulator in (2.1) can be reformulated as
ẋ1 = x2
ẋ2 = f x + dx + γx (6.1)
ż 1 = z 2
ż 2 = f z + dz + γz (6.2)
where x1 , z 1 denote the joint angles q1 and q2 , respectively,
[ f x , f z ]T = −M(q)−1 (C(q, q̇)q̇ + D q̇ + G(q, q b ))
are known functions of the states, [dx , dz ]T = −M(q)−1 H (q, q̇, q̇ b , q̈ b ) are
unknown disturbances due to the base oscillation, and [γx , γz ]T = M(q)−1 τ are
the transformed control inputs.
The RSSIs are designed as

t
sx = ėx + ax (ex )ex + bx (ex ) ex dτ (6.3)
0
t
sz = ėz + az (ez )ez + bz (ez ) ez dτ (6.4)
0
where a, b denote the proportional and integral control gains, respectively, ex =

x1 − r1 , ez = z 1 − r2 are the tracking errors, r1 , r2 are the joint angle references to
be tracked as mentioned before, and t is the time variable. Note that variables with
suffixes x, z denote ones associated with Links 1 and 2, respectively. a and b vary
with the tracking error as follows.
δax
ax (ex ) = axmin + (6.5)
cosh(erax )
δaz
az (ez ) = azmin + (6.6)
cosh(eraz )
δbx
bx (ex ) = bxmin + (6.7)
cosh(erbx )
δbz
bz (ez ) = bzmin + (6.8)
cosh(erbz )
where
δax = axmax − axmin (6.9)

δaz = azmax − azmin (6.10)
δbx = bxmax − bxmin (6.11)
δbz = bzmax − bzmin (6.12)
ex
erax = (6.13)
σax
ez
eraz = (6.14)
σaz
ex
erbx = (6.15)
σbx
ez
erbz = , (6.16)
σbz
the parameters with min and max are constants to determine the corresponding
minimal and maximal gains, respectively, and σa , σb are constants to adjust the
corresponding gain varying rates.
Remark 6.1 SMC-RSSI systems contain SMC-I ones in a natural manner such that
the minimal and maximal gains are set to the same. Further, by setting δa and δb
6.3 Sliding-Mode Control via Rotating Sliding Surface … 101
to small enough, and/or σ to large enough, the SMC-RSSI system can be arbitrarily
close to the corresponding SMC-I one.
To constrain the dynamics to the RSSIs sx = 0 and sz = 0, control inputs are

designed as
γx = vx − k x sgn(sx ) (6.17)
γz = vz − k z sgn(sz ) (6.18)
δax sinh(erax )
vx = − f x − ax ėx + ex ėx
σax cosh2 (erax )
t
δbx sinh(erbx )
+bx ex + ėx ex dτ (6.19)
σbx cosh (erbx ) 0
2
δaz sinh(eraz )
vz = − f z − az ėz + ez ėz
σaz cosh2 (eraz )
t
δbz sinh(erbz )
+bz ez + ėz ez dτ (6.20)
σbz cosh (erbz ) 0
2
where k x , k z are positive constant gains such that k x > |dx − r̈1 | and k z > |dz − r̈2 |.
Then, we analyze the sliding condition using the example of link 1. Note that the
case of link 2 is exactly the same as that of link 1. Let
1 2
Vs (sx ) = s (6.21)
2 x
be a Lyapunov function candidate for the sliding condition. Differentiating Vs (s)
with respect to t and substituting (6.1) and (6.17) yields
V̇s (sx ) = sx ṡx

t

= sx ëx + ax ėx + ȧx ex + bx ex + ḃx ex dτ
0
t

= sx ẋ2 − r̈1 + ax ėx + ȧx ex + bx ex + ḃx ex dτ
0

= sx f x + dx + vx − k x sgn(sx ) − r̈1 + ax ėx
δax sinh(erax )
+ ex ėx + bx ex
σax cosh2 (erax )
t
δbx sinh(erbx )
+ėx ex dτ
σbx cosh (erbx ) 0
2
= sx {dx − r̈1 − k x sgn(sx )}

≤ −(k x − |dx − r̈1 |)|sx |
< 0 (for sx = 0) (6.22)
since k x > |dx − r̈1 |. Thus, the sliding condition is satisfied.

Next, we provide a perspective on how to choose the gain parameters of RSSI in

(6.9)–(6.16). First, let us focus on the steady-state performance which is determined
by the maximal gains. In this study, we assume that the references r1 and r2 are to
be step, sinusoidal signals, and their linear sums. Thus, the integral gain b needs to
be large so that the bandwidth is large enough for the given references. However,
this should be generally done by taking into account of some constraints of practical
implementation, i.e., the digital control sampling period, actuator limitations, and
sensor noises. Then, the proportional gain a should be chosen under consideration
of an appropriate balance with the b. Meanwhile, the transient performance can be
modified by the minimal gains and gain varying rates. By properly choosing those
parameters, the transient performance can be improved in comparison with the SMC-
I controller, which is demonstrated later by simulations.
On the other hand, k x and k z in (6.17) and (6.18) are the most important para-
meters in order to constrain the dynamics to the sliding surface as proven above.
Theoretically and without constrains of practical implementation, the lager k x and
k z lead to the stronger robustness against disturbances and model uncertainties which
satisfy the matching condition. Their appropriate values depend on the problems and
situations. As demonstrated by simulations, appropriate values of k x and k z in the
nominal-case problem and in the robust control problem are largely different, because
in the robust control problem dx and dz contain not only the disturbance due to the
base oscillation but also the error of canceling the nonlinearity f x and f z .
6.3.2 Control System Design Example
Table 6.1 shows controller parameters of an example of SMC-RSSI which is utilized

to demonstrate stability analysis of RSSI in the next section. Using this SMC-RSSI,
simulations of attitude control in local coordinates in the case of single-frequency
oscillation were conducted. In Fig. 6.1, the upper graph shows the result with the
sampling period of 0.01 s while the lower one does that with the sampling period
of 0.001 s. The simulation conditions are the same as those for the H∞ controllers
in Chap. 5 except the sampling time. First, the SMC-RSSI reveals successful per-
formances in the both sampling period cases. On the other hand, we see that the
controller is largely influenced by the sampling period, which is the characteristic
feature of SMC. This fact is natural because in theory the SMC scheme assumes an
infinite velocity switching control inputs.
Table 6.1 Controller parameters of an example of SMC-RSSI

Link amin amax σa bmin bmax σb k
1 5 40 0.7 80 350 0.7 6
2 5 80 0.7 100 700 0.7 4
6.4 Stability Analysis of RSSI 103
100
Link 1
Link 2
50
qi (deg)
0 2 4 6 8 10 12 14 16 18 20
100
Link 1
Link 2
50
qi (deg)
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.1 Simulation results of attitude control in local coordinates with the SMC-RSSI example for
the single-frequency oscillation; top with the sampling period of 0.01 s, bottom with the sampling
period of 0.001 s
6.4 Stability Analysis of RSSI
The time-varying sliding surface in [11, 12] and the nonlinear sliding surface in [32]
are scalar systems and hence their stabilities are guaranteed as long as their slopes
are negative. However, the RSSI is a nonlinear second-order system and the problem
of its stability is nontrivial. Therefore, we need stability analysis of RSSI to confirm
that the RSSI is a stable manifold.
In this monograph, we analyze two cases of global stability and semi-global sta-
bility, respectively. To analyze stability of RSSI, we prepare a lemma which provides
a sufficient condition to construct a Lyapunov function. If this condition is globally
satisfied on the RSSI, then the origin of the RSSI is globally asymptotically stable.
Otherwise its semi-global stability can be proved by conducting the following
procedures. First, we differentiate regions on the RSSI where the condition in the
lemma is satisfied and not satisfied, respectively. Second, by utilizing the Lyapunov
function and a numerical solution, we show that there exists an invariant compact set
containing the origin as a unique equilibrium for trajectories on the RSSI. Finally,
by showing that any closed orbits, that is, periodical solutions do not exist in the
invariant compact set, it is proved that the RSSI is asymptotically stable with respect
to the origin within the set. Respective theorems regarding the global stability and
semi-global stability of RSSI are given.
Here, we prepare the following lemma which provides a machinery to construct
a Lyapunov function for a time-varying and/or nonlinear second-order system.
Lemma 6.1 Consider the following second-order system,

˙
ie ie
= Ae
ė e

0 1 ie
= (6.23)
−(b + δb) −(a + δa) e
where e = [ie, e]T ∈ D is the state vector, a > 0 and b > 0 are real constants,
and δa, δb are real functions of time and/or the state variables. D is the real domain
which contains the origin. Suppose the matrix

0 1
Ae0 = (6.24)
−b −a
has a pair of complex conjugate eigenvalues −α ± iβ, α > 0, β > 0 ∈ R. Then,

using the matrix

1 1
Se = , (6.25)
−α + β −α − β
V (e ) = e (Se−1 )T Se−1 e

T
(6.26)
can be a Lyapunov function for the system in (6.23) with δa = 0 and δb = 0. Further,
let δ p = δb − αδa, and if
1
δa > δ p2 − α (6.27)
4αβ 2
over D, then V (e ) is also a Lyapunov function for the system in (6.23).
Proof Since |Se | = −2β = 0, Se is nonsingular and hence can be a similarity

transformation to the system in (6.23). Then, Ae is transformed into
Pe = Se−1 Ae Se .
For Ae0 we use Pe0 accordingly.

Further, V (e ) is positive definite. Thus, differentiating V (e ) with respect to time
yields
V̇ (e ) = e (Se−1 )T (PeT + Pe )(Se−1 )e .

T
First, we consider the case where δa = 0 and δb = 0. Then,

−2α 0
Pe0 + Pe0 =
T
≺ 0,
0 −2α
which proves that V̇ (e ) is negative definite, together with the nonsingularity of Se .
Thus, V (e ) is a Lyapunov function.
Next, we investigate the case where δa = 0 and/or δb = 0. In this case, by some
algebraic manipulations with
a = 2α,
b = α2 + β 2 ,
we obtain

1 2αβ + δa(−α + β) + δb −δaβ
PeT + Pe = − .
β −δaβ 2αβ + δa(α + β) − δb
Being based on Sylvester’s criterion, this matrix is negative definite, if and only if
the following inequalities hold
1
δa > − δ p − 2α,
β
1
δa > − δ p 2 − α,
4αβ 2
where δ p = δb − αδa. However, since
1 1 1
{− δ p 2 − α} − {− δ p − 2α} = (δ p + 2αβ)2 ≥ 0
4αβ 2 β 4αβ
those conditions can be integrated into the inequality (6.27). Therefore, if (6.27)
holds, then V (e ) is a Lyapunov function for the system in (6.23), and the proof is
completed.
Note that e and ie correspond to the tracking error and its integral on the RSSI,
and that the complex eigenvalue condition is the case for the RSSI to be analyzed.
Now, we demonstrate two cases of stability analysis by deploying the example of
SMC-RSSI in Table 6.1. First, we derive the following theorem regarding the global
stability of RSSI for the case of Link 1 with the SMC-RSSI.
Theorem 6.1 (Global stability) Consider the RSSI represented by the following
second-order system.

˙
ie 0 1 ie
= (6.28)
ė −b(e) −a(e) e
where the domain D of the state vector e = [ie, e]T is R × [−π, π], a(e) = ax (ex )
in (6.5), b(e) = bx (ex ) in (6.7), and the associated parameters are defined as those
for Link 1 with the SMC-RSSI in Table 6.1. Then, the RSSI is globally asymptotically
stable with respect to the origin.
Link 1
40
e =0
30
20
δa
10
e =π
0
−10
−200 −150 −100 −50 0 50 100 150 200
δp
Link 2
80
e =0
60
e =0.32159
40
δa
20
e =π
0
−100 −50 0 50 100 150 200 250 300 350 400

δp
Fig. 6.2 Results of checking the condition for a Lyapunov function in (6.27) for the SMC-RSSI
in Table 6.1; solid line boundary for the condition in (6.27), dashed line parameter trajectory with
respect to e, top Link 1, bottom Link 2
Proof Since it can be straightforwardly confirmed that the right-hand side function
in (6.28) is globally Lipschitz in e , the existence and uniqueness of the solution to
every initial point e (0) ∈ D is ensured. Further, due to the nonsingularity of the
right-hand side matrix, the origin is the unique equilibrium.
Then, we construct a Lyapunov function candidate with e = π and check the
sufficient condition in (6.27) with Lemma 6.1. The result is shown in the top graph in
Fig. 6.2, where the dashed line represents (δ p, δa) over e ∈ [0, π] and the solid line
does the boundary by (6.27). Note that a(·) and b(·) are even functions with respect
to e. Thus, it is seen that the condition in (6.27) is globally satisfied on the domain
D. Therefore, the proof is completed.
Next, we present a semi-global stability case with the example of Link 2 as in the
following theorem.
Theorem 6.2 (Semi-global stability) As in Theorem 6.1, consider the RSSI repre-
sented by (6.28) over the same domain and the corresponding parameters are defined
as those for Link 2 with the SMC-RSSI in Table 6.1. Then, the RSSI is semi-globally
asymptotically stable with respect to the origin.
Proof The existence and uniqueness of the solution and the property of the origin on
the RSSI are the same as in Theorem 6.1. We build a Lyapunov function candidate
with e = π and check the condition by (6.27). The bottom graph in Fig. 6.2 shows
the result. It is seen that the condition is satisfied for 0.32159 ≤ |e| ≤ π, however is
not for 0 ≤ |e| < 0.32159.
Ev
2 Region 1
e 0 Region 2
−2 Region 1
−0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4

ie
0.5 e =0.32159
P
1
0
e
Trajectory from P
1
e =−0.32159
P P
−0.5 2 3
0.3 0.31 0.32 0.33 0.34 0.35 0.36 0.37

ie
Fig. 6.3 Invariant compact set on the phase plane; top on the normal scale, bottom zoomed-in area
encircled by the dotted rectangle in the top figure
Then, we show that there exists an invariant compact set of the RSSI. To this end,
we define regions on the domain D as {ie, e|ie ∈ R, 0.32159 ≤ |e| ≤ π} (Region
1) where the condition in (6.27) is satisfied and {ie, e|ie ∈ R, 0 ≤ |e| ≤ 0.32159}
(Region 2) where the condition is not satisfied.
Next, we define V (e ) = const such that the corresponding ellipsoid Ev on the
ie-e phase plane passes through (0, π), which is depicted in Fig. 6.3 (see the top one).
Focus on the area encircled by the dotted rectangle which is magnified on the bottom
figure in Fig. 6.3. By calculating the solution with the initial point P1 , we can obtain
the trajectory from P1 which passes through Region 2 and goes into Region 1 inside
Ev. On the other hand, the trajectory from every point on the line segment P2 P3 also
goes into Region 1 inside Ev according to the vector field. Considering those facts
and the uniqueness of the solution together with the symmetry of the vector field by
(6.28) with respect to the origin, therefore, it is concluded that the compact set with
the boundary consisting of the parts of Ev in Region 1, the trajectory from P1 to P2 ,
the line segment P2 P3 and their symmetric counterparts is an invariant set. That is,
the trajectory from every point in the set is constrained in the set.
Hence, by applying Poincare-Bendixson theorem, if there exists no periodic solu-
tion in this compact set, it is ensured that the compact set is asymptotically stable
with the origin. We show which by contradiction.
Suppose that there exists a periodic solution e (t) = e (t + T ) in the compact set,
and define the energy function
1 1
E(e ) = b(e)ie2 + e2 . (6.29)
2 2
Then,
0 = E(e (t + T )) − E(e (t))

t+T
= Ėdτ
t
t+T
= ˙ + eė + 1 ḃie2 }dτ
{b(e)ieie
t 2
t+T 1
= {b(e)iee + e(−b(e)ie − a(e)e) + ḃie2 }dτ
t 2
t+T t+T
1 db de
=− a(e)e2 dτ + ie2 dτ
t 2 t de dτ
t+T
1 e(t+T ) 2 db
=− a(e)e2 dτ + ie de
t 2 e(t) de
t+T
=− a(e)e2 dτ
t
<0 (6.30)
which is contradiction. Thus, the proof is completed.
Note that this asymptotically stable region is sufficiently large for a practical
application.
6.5 Simulations
In order to evaluate the proposed SMC-RSSI systems, control simulations were

performed emphasizing comparison with the proposed H∞ control systems. The
demonstration cases are the same as in the case of H control in Chap. 5. First, for the
nominal case, i.e., the payload width is 5 mm, all the combinations of the three types
of control problems and three types of base oscillation will be presented. Then, robust
control demonstrations will be addressed. In all the demonstrations, the methods and
conditions are almost the same as those presented in Chap. 5 except the sampling
period and the sensor resolution. As illustrated in Fig. 6.1, SMC systems are largely
influenced by the sampling period. Even in the nominal case, in order that the SMC-
RSSI performs satisfactorily for those various demonstration cases, the sampling
period of 0.001 s is required. Moreover, the control system is also affected by the
sensor resolution. Two sensor resolutions of 0.18 and 0.0018 deg are compared for the
simulations, to investigate such properties of the system and show that how effective
the system can be under desirable conditions of sensor resolution. For the simulation
cases, the control parameters were tuned through a trial and error manner, which are
shown in Table 6.2.
6.5 Simulations 109
Table 6.2 Controller parameters of SMC-RSSIs for simulation cases

Case Link amin amax σa bmin bmax σb k
Nominal 1 25 25 – 1 4000 0.3 6
2 30 30 – 1 1500 0.7 4
Robust 1 25 25 – 1 4000 0.3 12
2 30 30 – 1 1500 0.7 45
6.5.1 Simulations in the Nominal Case
Table 6.3 compares the simulation results of the SMC-RSSI with 0.18 deg sensor
resolution, the SMC-RSSI with 0.0018 deg one, the H∞ controllers with 0.18 deg
one, and the H∞ controllers with 0.0018 deg one, with respect to RMSE ē. Note
that those controllers were performed with the sampling period of 0.001 s except the
H∞ controllers with 0.18 deg sensor resolution (with that of 0.01 s).
First, let us compare the results of the SMC-RSSI and H∞ controllers in the case
of sensor resolution of 0.18 deg. Except the position control cases, the H∞ controllers
performed slightly better the SMC-RSSI. Further, it should be noted that this SMC-
RSSI requires five times larger toque limits than those for the other controllers.
Then, we review the results in the case of finer sensor resolution of 0.0018 deg.
Interestingly enough, the results of the SMC-RSSI has been considerably improved
with the finer sensor resolution. Similarly, those of the H∞ controllers has been
Table 6.3 Results of RMSE ē in the respective control simulations

Single SMC-RSSI (0.18 SMC-RSSI K 21 (0.18 deg) K 21 (0.0018 deg)
deg) (0.0018 deg)
Attitude-l (deg) 0.2088 0.1643 0.1047 0.0557
Attitude-g (deg) 0.2131 0.1500 0.2152 0.2034
Position (mm) 0.2336 0.0559 0.3972 0.3846
Double SMC-RSSI (0.18 SMC-RSSI K 22 (0.18 deg) K 22 (0.0018 deg)
deg) (0.0018 deg)
Attitude-l (deg) 0.1690 0.0318 0.1072 0.0630
Attitude-g (deg) 0.1775 0.0346 0.1361 0.0721
Position (mm) 0.2073 0.0137 0.2276 0.1410
Bretschneider SMC-RSSI (0.18 SMC-RSSI K 23 (0.18 deg) K 23 (0.0018 deg)
deg) (0.0018 deg)
Attitude-l (deg) 0.1773 0.0046 0.1085 0.0163
Attitude-g (deg) 0.1988 0.0056 0.1486 0.0401
Position (mm) 0.2712 0.0133 0.2358 0.1283
Attitude-l attitude control in local coordinates, attitude-g attitude control in global coordinates,
position position control. Note that the SMC-RSSI with 0.18 deg resolution requires five times
larger torque amplitude limits
100
Link 1
qi (deg)
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.4 Simulation results of attitude control in local coordinates with the SMC-RSSI (0.18 deg)
for the double-frequency oscillation; top the attitudes of the links in the local coordinates, second
from the top: the tracking errors on a normal scale, third from the top: the tracking error of Link 1
improved, but the degree of improvement is less than that of the SMC-RSSI. In
other words, the SMC-RSSI tends to be influenced by sensor resolution and noise in
terms of control performance and required control inputs, being compared to the H∞
controllers. On the other hand we should note that for various types of base oscillation
the single SMC-RSSI can work well, whereas the corresponding weighting functions
to the assumed base oscillation should be prepared to design an H∞ controller.
Next, we present the results in time-series graphs by choosing examples. The
double-frequency demonstrations have been selected to display the time-series pro-
files of control performance. For each control problem, the control performances of
the SMC-RSSIs with the sensor resolutions of 0.18 and 0.0018 deg are compared.
Figures 6.4 and 6.5 present the simulation results of attitude control in local coor-
dinates with the respective SMC-RSSIs. Then, Figs. 6.6 and 6.7 show the results of
attitude control in global coordinates, and Figs. 6.8 and 6.9 show those of position
control. As seen from the results, both the SMC-RSSIs suppress the disturbances and
achieve successful control performance. In particular, by comparing the control per-
formances with those of the H∞ controllers, not only the steady-state performances
but also the transient performances at just after 10 s are considerably successful,
which is exactly the benefit of the RSSI as demonstrated later. Furthermore, it is seen
that such a finer sensor resolution can dramatically improve the control performance
of the SMC-RSSI.
By changing the view point, we examine profiles of control inputs of the SMC-
RSSIs which represent the characteristic feature of SMC schemes. The results of
6.5 Simulations 111
100
Link 1
qi (deg)
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.5 Simulation results of attitude control in local coordinates with the SMC-RSSI (0.0018
deg) for the double-frequency oscillation; top the attitudes of the links in the local coordinates,
second from the top: the tracking errors on a normal scale, third from the top: the tracking error of
Link 1 on a fine scale, bottom the tracking error of Link 2 on a fine scale
100
qgi (deg)
Link 1
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.6 Simulation results of attitude control in global coordinates with the SMC-RSSI (0.18 deg)
for the double-frequency oscillation; top the attitudes of the links in the local coordinates, second
from the top: the tracking errors on a normal scale, third from the top: the tracking error of Link 1
100
qgi (deg)
Link 1
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
5
error (deg)
Link 1
0 Link 2
−5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 1
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
error (deg)
0.5 Link 2
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.7 Simulation results of attitude control in global coordinates with the SMC-RSSI (0.0018
deg) for the double-frequency oscillation; top the attitudes of the links in the local coordinates,
second from the top: the tracking errors on a normal scale, third from the top: the tracking error of
Link 1 on a fine scale, bottom the tracking error of Link 2 on a fine scale
150
Link 1
qgi (deg)
100
Link 2
50
0
0 2 4 6 8 10 12 14 16 18 20
0.1
error (m)
0.05
0
0 −3 2 4 6 8 10 12 14 16 18 20
x 10
5
error (m)
0
0 2 4 6 8 10 12 14 16 18 20
torque (Nm)
0.5
Link 1
0 Link 2
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.8 Position control simulation results with the SMC-RSSI (0.18 deg) for the double-frequency
oscillation; top the attitudes of the links in the global coordinates, second from the top: the distance
errors from the final goal on a normal scale, third from the top: the distance errors from the final
6.5 Simulations 113
200
Link 1
0 Link 2
−200
0 2 4 6 8 10 12 14 16 18 20
0.1
error (m)
0.05
0
0 −3
2 4 6 8 10 12 14 16 18 20
x 10
5
error (m)
0
0 2 4 6 8 10 12 14 16 18 20
torque (Nm)
0.1
Link 1
0 Link 2
−0.1
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.9 Position control simulation results with the SMC-RSSI (0.0018 deg) for the double-
frequency oscillation; top the attitudes of the links in the global coordinates, second from the top:
the distance errors from the final goal on a normal scale, third from the top: the distance errors from
the final goal on a fine scale, bottom the control torques
control toque inputs in the simulations of the respective control problems, links,
and SMC-RSSIs are shown in Figs. 6.10, 6.11, 6.12, 6.13, respectively. First, those
profiles exhibit features like thick belts when compared to those of the H∞ controller
which seems to be thin strings (see Figs. 5.14 and 5.15 again), which is exactly the
characteristic feature of switching control by SMC. Then, comparing those of the
SMC-RSSIs with 0.18 and 0.0018 deg, Figs. 6.10 and 6.11 (case of 0.18 deg) reveal
five times larger torques than those in Figs. 6.12 and 6.13 (case of 0.0018 deg). In
particular, in the case of 0.18 deg, the required switching amplitude might be tough
for practical implementation to actuators. We should note that this characteristic
feature of control input give rise to the serious gap between the theory and practical
application of the SMC scheme, such as chartering phenomena, unexpected control
errors.
Here, we address effectiveness of the RSSI being compared to the conventional
constant-gain sliding surface. To this end, we performed step-tracking attitude control
in local coordinates for the double-frequency oscillation for the SMC-RSSI and a
SMC-I without any low-pass filter for the step signal, so that the effectiveness of the
RSSI becomes apparent. The SMC-I employed the maximal gains of the SMC-RSSI
(in the nominal case in Table 6.2) as its constant gains. First, Fig. 6.14 compares
their results in the case of unbounded control torques. As seen from the figure, the
both results exhibit satisfactory steady-state performances similarly while the SMC-I
system suffers from large overshoot during the transient state. Figure 6.15 display
0.5
torque (Nm)
0
−0.5
0 2 4 6 8 10 12 14 16 18 20
0.5
torque (Nm)
−0.5
0 2 4 6 8 10 12 14 16 18 20
0.5
torque (Nm)
−0.5
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.10 Control torques of the joint 1 with the SMC-RSSI (0.18 deg) for the respective problems
in the case of double-frequency oscillation; top attitude control in local coordinates, middle attitude
control in global coordinates, bottom position control in global coordinates
0.1
torque (Nm)
−0.1
0 2 4 6 8 10 12 14 16 18 20
0.1
torque (Nm)
−0.1
0 2 4 6 8 10 12 14 16 18 20
0.1
torque (Nm)
−0.1
0 2 4 6 8 10 12 14 16 18 20
time (s)
6.5 Simulations 115
0.1
torque (Nm)
0
−0.1
0 2 4 6 8 10 12 14 16 18 20
0.1
torque (Nm)
−0.1
0 2 4 6 8 10 12 14 16 18 20
0.1
torque (Nm)
−0.1
0 2 4 6 8 10 12 14 16 18 20
time (s)
0.02
torque (Nm)
−0.02
0 2 4 6 8 10 12 14 16 18 20
0.02
torque (Nm)
−0.02
0 2 4 6 8 10 12 14 16 18 20
0.02
torque (Nm)
−0.02
0 2 4 6 8 10 12 14 16 18 20
time (s)
100
Link 1
Link 2
50
qi (deg)
0 2 4 6 8 10 12 14 16 18 20
100
Link 1
Link 2
50
qi (deg)
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.14 Comparison of simulation results with the SMC-RSSI and SMC-I in the case of
unbounded control torque (attitude control in local coordinates for the double-frequency oscil-
lation); top SMC-RSSI, bottom SMC-I
4000 Link 1
Link 2
3000
b
2000
1000
0
0 2 4 6 8 10 12 14 16 18 20
time (s)
1
Link 1
0.8 Link 2
0.6
ζ
0.4
0.2
0
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.15 Variation of bx , bz , and the damping ratios of simulation results with the SMC-RSSI
and SMC-I in the case of unbounded control torque (attitude control in local coordinates for the
double-frequency oscillation); top SMC-RSSI, bottom SMC-I
6.5 Simulations 117
100
Link 1
Link 2
50
qi (deg)
0 2 4 6 8 10 12 14 16 18 20
100
Link 1
Link 2
50
qi (deg)
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.16 Comparison of simulation results with the SMC-RSSI and SMC-I in the case of bounded
control torque (attitude control in local coordinates for the double-frequency oscillation); top SMC-
RSSI, bottom SMC-I
how the SMC-RSSI varies the proportional gains bx and bz and correspondingly the
damping ratios in the control simulation. The gains decrease steeply and the damping
ratios increase accordingly, which plays a key role in suppressing the overshoot and
control inputs.
The point is that when the tracking error is large, the controller can accelerate the
object enough even with a small proportional gain and even tries to suppress too large
velocity and undesirable large surge control input, then as the error is getting small
the controller tries to keep a moderate velocity by increasing the proportional gain
accordingly within the range which will not lead to overshooting. Which enables
the controller to achieve such a rapid response in control performance without over-
shooting and too large control inputs.
Next, we examine the results of the same simulation in the case of bounded control
torques similarly to that of the H∞ controllers and most of the SMC-RSSIs men-
tioned before. Figure 6.16 presents the results, where the SMC-RSSI reveals a similar
successful performance to that in the case of unbounded control torque, but with a
slightly slower response, whereas the SMC-I does unacceptably bad transient-state
performance. Therefore, all the above arguments have proved that the RSSI can pro-
vide effectiveness in improving transient-state control performance and suppressing
control inputs.
Finally, in this section, we demonstrate how the SMC-RSSI is largely influenced
by the sensor resolution and sampling period. Figure 6.17 shows the simulation result
of attitude control in global coordinates for the double-frequency oscillation with the
SMC-RSSI with the sensor resolution of 0.18 deg and the sampling period of 0.01s,
100
Link 1
qgi (deg)
50 Link 2
0
0 2 4 6 8 10 12 14 16 18 20
20
error (deg)
Link 1
0 Link 2
−20
0 2 4 6 8 10 12 14 16 18 20
10
error (deg)
Link 1
0
−10
0 2 4 6 8 10 12 14 16 18 20
10
error (deg)
Link 2
0
−10
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.17 Simulation results of attitude control in global coordinates with the SMC-RSSI (0.18
deg) with the sampling period of 0.01 s for the double-frequency oscillation; top the attitudes of the
links in the local coordinates, second from the top: the tracking errors on a normal scale, third from
the top: the tracking error of Link 1 on a fine scale, bottom the tracking error of Link 2 on a fine
scale
where it is seen that such a condition deteriorate the control performance of the
SMC-RSSI seriously. Which is also one of features representing the gap between the
theory and practical application of the SMC scheme.
6.5.2 Simulations in the Robust Control Case
In this section, we present simulation results in the robust control case, where all the
combinations of the three types control problems and three types of base oscillations
with the payload variation with the widths, 0, 1, 2, . . . , 9, 10 mm were conducted.
Not like in the nominal-case demonstrations, we did not use the SMC-RSSI with
the sensor resolution of 0.18 because the controller cannot achieve acceptable roust
control performance. Moreover, since the SMC-RSSI (0.0018 deg, k x = 6, k z = 4)
cannot satisfactory robust control performance compared to those of the H∞ con-
trollers, we prepared a new SMC-RSSI of which robustness has been enhanced by
increasing k x to be 12 and k z to be 45. As shown in Table 6.2, the other parameters
are the same.
As in the case of the H∞ controllers, the simulation results are expressed being
based on the RMSE ē’s versus the payload widths in the respective control problems.
Figures 6.18 and 6.19 show the results of the robust control simulations with the
6.5 Simulations 119
1.5
RMSE (deg)
1
0.5
0
0 1 2 3 4 5 6 7 8 9 10
1.5
RMSE (deg)
0.5
0
0 1 2 3 4 5 6 7 8 9 10
2
RMSE (mm)
0
0 1 2 3 4 5 6 7 8 9 10
payload width (mm)
Fig. 6.18 RMSE ē versus payload width of the simulation results of robust control with the SMC-
RSSI (0.0018 deg, k x = 6, k z = 4); top attitude control in local coordinates, middle attitude control
in global coordinates, bottom position control in global coordinates; solid single frequency, dashed
double frequency, dashed-dotted Bretschneider
0.2
RMSE (deg)
0.1
0
0 1 2 3 4 5 6 7 8 9 10
0.2
RMSE (deg)
0.1
0
0 1 2 3 4 5 6 7 8 9 10
0.2
RMSE (mm)
0.1
0
0 1 2 3 4 5 6 7 8 9 10
payload width (mm)
Fig. 6.19 RMSE ē versus payload width of the simulation results of robust control with the SMC-
RSSI (0.0018 deg, k x = 12, k z = 45); top attitude control in local coordinates, middle attitude
control in global coordinates, bottom position control in global coordinates; solid single frequency,
dashed double frequency, dashed-dotted Bretschneider
SMC-RSSI (0.0018 deg, k x = 6, k z = 4) and SMC-RSSI (0.0018 deg, k x = 12,

k z = 45), respectively. To simply differentiate those controllers, we refer to the
former controller as SMC-RSSI1 and the latter one as SMC-RSSI2. The figures have
graphs which show the single-frequency case, the double-frequency case, and the
Bretschneider case, from the top to the bottom, respectively. In each graph, the solid
line represents the results of attitude control in local coordinates, the dashed line
does those of attitude control in global coordinates, and the dashed-dotted line does
those of position control. As mentioned above, the control performances of the SMC-
RSSI1 are considerably deteriorated by the payload variation as shown in Fig. 6.18,
being compared to those of the H∞ controllers. Further, it is interesting to notice
that, regardless of control problems and base-oscillation patterns, the relationships
between RMSE variation and payload variation are very similar, which is almost
linear and symmetric with respect to the nominal payload with width of 5 mm. This
feature is very unique regarding the profiles of the other controllers (see Figs. 5.28–
5.31 again). For instance, the H∞ controllers without explicit consideration of model
uncertainties and the PID one reveal the feature that RMSE decreases with payload
increase.
On the other hand, the control performances of the SMC-RSSI2 enhanced with
respect to robustness is extremely superior to those of the other controllers as seen
in Fig. 6.19. It should be noted that the order of RMSEs the SMC-RSSI2 is less than
those of the other controllers by a factor of 10. This result demonstrates that how easily
one can increase the robustness of such a SMC-based controller by only increasing the
0.1
torque (Nm)
−0.1
0 2 4 6 8 10 12 14 16 18 20
0.1
torque (Nm)
−0.1
0 2 4 6 8 10 12 14 16 18 20
0.1
torque (Nm)
−0.1
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.20 Control torques of the joint 1 with the SMC-RSSI (0.0018 deg, k x = 12, k z = 45) for
the respective problems in the case of double-frequency oscillation with the nominal payload of 5
mm; top attitude control in local coordinates, middle attitude control in global coordinates, bottom
position control in global coordinates
6.5 Simulations 121
0.02
torque (Nm)
0
−0.02
0 2 4 6 8 10 12 14 16 18 20
0.02
torque (Nm)
−0.02
0 2 4 6 8 10 12 14 16 18 20
0.02
torque (Nm)
−0.02
0 2 4 6 8 10 12 14 16 18 20
time (s)
Fig. 6.21 Control torques of the joint 2 with the SMC-RSSI (0.0018 deg, k x = 12, k z = 45) for
the respective problems in the case of double-frequency oscillation with the nominal payload of 5
mm; top attitude control in local coordinates, middle attitude control in global coordinates, bottom
position control in global coordinates
switching control amplitude as k x and k z , which is the most advantageous property

of the SMC schemes. However, the larger switching control amplitude means the
more difficulty in its implementation in practice. Then, let us see the control torque
inputs of the SMC-RSSI2 in the nominal payload case in Figs. 6.20 and 6.21. By
comparing the results to those of the SMC-RSSI1 in Figs. 6.12 and 6.13, it is seen
that the extremely large amplitude is required for the SMC-RSSI2 according to the
amplitude of k x and k z . Therefore, practical implementation of the SMC-RSSI2
might lead to undesirable chattering phenomena and/or unexpected control error due
to the mismatching capability of actuators to the theory.
6.6 Conclusions
In this chapter, we have explained our proposed SMC-based control approach, i.e., the
SMC-RSSI, conducted stability analyses with respect to the RSSI, and demonstrated
its capabilities by simulations. In order to compare this approach to the proposed
H∞ -control-based one in the foregoing chapters, the method for demonstration is
consistent with that for the H∞ -control-based one. Consequently, we have concluded
that, with respect to comparison to the H∞ -control-based approach:
• Without the knowledge of the frequency-domain property of base oscillation, an

SMC-RSSI can be designed and the single controller can appropriately perform
for various types of base oscillations;
• The SMC-RSSI can provide successful control performances not only in the steady
state, but also in the transient state without any additional control structure such
as a TDOF control structure;
• By only increasing the switching control amplitude, one can easily enhance the
robustness of controller;
• However, the SMC-RSSI is considerably sensitive to the control system conditions
such as sensor resolution and noise, sampling period, actuator capability, and so
on. Hence, when implemented in practice, any mismatches of such conditions
of the practical control system to the theory will lead to serious problems such
as undesirable chattering phenomena, unexpected control error, too large control
inputs.
Moreover, comparing the proposed approach to the conventional SMC, we have
confirmed that
• The SMC-RSSI can achieve the superior transient control performance such as
rapid response without overshooting, and can suppress large control input.
In this monograph, we have not presented experimental demonstration for the SMC-
based approach, because the present experimental OBM is not capable of appro-
priately implementing the SMC-RSSI in terms of the sensor resolution, sampling
period, switching ability of the actuators. This point is one the next steps to be taken
in future works.
Chapter 7
Base Oscillation Estimation via Multiple H∞
Filters
Abstract This chapter presents a complementary technique to support the afore-

mentioned control methodologies which have assumed that accurate measurements
of the base oscillation can be available for the control system. Here we present an
estimation algorithm of the base oscillation assuming a low-cost rate gyro sensor.
The heart of this algorithm is a method of selectively combining multiple H∞ filters
according to an innovation-based criterion. We introduce this algorithm and demon-
strate its estimation performance by simulations using the Bretschneider oscillation
in Chap. 2. The results show that among the multiple filters appropriate ones can
be selected with the innovation-based criterion, and thus the estimation algorithm is
effective.
7.1 Introduction
In the foregoing chapters, both the proposed control methods have been addressed
and demonstrated assuming that accurate measurements of the base oscillation are
available. In practice, the accuracy of measurements depends on sensing systems
used. In general, a high-precision sensing system such as a fiber optical gyro-based
system requires expensive cost. Therefore, we have been involved in development
of ship oscillation estimation algorithm using a relatively low-cost rate gyro sensor
in order to support and enhance the assumption for the control methods.
In this chapter, an estimation algorithm of ship oscillation with a low-cost rate
gyro sensor is presented. In particular, situations in the presence of frequency-domain
property variation of oscillation are considered, which necessarily occur due to
changing the ship velocity and direction, i.e., the way of encountering the ocean
waves.
There have been typical two research directions to deal with motion estimation
methods for moving objects. One of which is to combine measurements from the
multiple sensors, so-called “sensor fusion technique” [55, 81], and the other one is
to develop or improve model-based filtering algorithm such as the Kalman filter [23,
29, 54, 63, 66, 100, 107]. As examples of the former works, Leavitt et al. [55] and
Sanca et al. [81] have proposed methods of compensating drift and integral errors
DOI 10.1007/978-3-319-21780-2_7
124 7 Base Oscillation Estimation via Multiple H∞ Filters
of rate gyro sensors using together other types of sensors, such as an accelerometer
and inclinometer for the purpose of developing an attitude estimation method. On
the other hand, as an example of the latter works, to predict wave-induced oscillatory
ship motions, Küchler et al. [54] have presented a motion estimation method for a
crane-vessel using a fast Fourier transform (FFT)-based state-equation model and
the conventional Kalman filter. Further, Liu et al. [63] have developed a method of
improving autoregressive model by the Kalman filter to estimate ship motions.
Motivated by [54] and adding a new viewpoint of variation of ship oscillation
frequency, we have developed an estimation method of ship oscillation via a low-
cost rate gyro sensor, which incorporates periodic updates of the FFT-based model
and a method of selectively combining multiple H∞ and Kalman filters, in order to
enhance adaptiveness and robustness of estimation for a model variation. In general,
such variation will deteriorate the estimation performance particularly when relying
on a model-based estimation algorithm such as the Kalman filter. The method of
combining multiple filters according to an innovation-based criterion is the key of
our estimation method. The Kalman filtering algorithm, which is one of the most
popular algorithms [23, 63, 100, 112], is to effectively reduce Gaussian white noises,
while the H∞ filtering one [33, 85], which has been becoming popular, e.g., [6, 27,
47, 49, 59, 78, 106], can accommodate norm-bounded deterministic noises, and
therefore is expected to be robust against model errors. Our method of combining
filters is to utilize the advantage of each filter as much as possible.
This chapter is organized as follows. In Sect. 7.2, we briefly explain the con-
ventional Kalman and H∞ filters and introduce our proposed method. Section 7.3
presents an ocean wave spectrum model and a ship motion model for demonstrations.
Then, Sect. 7.4 demonstrates estimation performances of our method by simulations.
The last section will give concluding remarks.
7.2 Estimation Algorithm
This section presents our proposed estimation algorithm. Figure 7.1 shows an outline
of the algorithm, which consists of the processes of acquisition of a measurement
of the angular velocity, periodic FFT-based model updates, selectively combining
H∞ and Kalman filters, integral with drift error compensation, and estimation of the
base inclination angle, TFFT is the FFT period, ΔtFFT is the interval to update the
FFT-based model.
How to build the FFT-based state-equation model in Fig. 7.1 is addressed in
Sect. 7.2.1. Then, Sects. 7.2.2 and 7.2.3 briefly review the conventional discrete
Kalman filter and H∞ filter, respectively. Next, Sect. 7.2.4 presents the method of
selectively combining multiple filters, which is the heart of our methodology. Finally,
to output the angle not velocity, the method of integral of velocity with drift error
compensation is explained in Sect. 7.2.5.
7.2 Estimation Algorithm 125
angular velocity measurement
no
t=T FFT +jΔt FFT ?
yes
Update of the FFT-based model
H∞ filter 1 H∞ filter 2 ••• H∞ filter N Kalman filter
Selectively combined
Integral with drift compensation
inclination angle estimation
Fig. 7.1 Outline of the estimation algorithm; 0 ≤ j ∈ Z: sampling index, TFFT : FFT period,
ΔtFFT : update period
7.2.1 FFT-Based Linear State-Equation Model
In accordance with the notations in the foregoing chapters, let qb (t) denote the base
inclination angle to be estimated, and thus q̇b (t) is the angular velocity to be measured
via a gyro sensor. Further, let q̇b (t) be decomposed into multiple sine waves called
“modes” using FFT. From all the modes, the n dominant modes with the n largest
amplitudes are chosen and then q̇b (t) is approximated by

n
q̇b (t) ≈ ai sin (ωi t + φi ) , (7.1)
i=1
where ai , ωi and φi are the amplitude, angular frequency, and phase of the ith mode,
respectively. Further, among the chosen modes, the angular frequency of the most
dominant mode, i.e., with the largest amplitude, is denoted by ωmax , which is utilized
in the processes of selectively combining multiple filters and integral with drift error
compensation as in the sequel.
The dominant modes are incorporated into a state-space equation as in the
following:
ẋ(t) = Ax(t), (7.2)

y(t) = C x(t) + v(t), (7.3)
⎡ ⎤
A1 0 · · · · · · 0
⎢ .. ⎥
⎢ 0 A2 . . . . ⎥
⎢ ⎥
⎢ .. . . . . . . .. ⎥
A= ⎢ . . . . . ⎥ ,
⎢ ⎥
⎢ . .. .. ⎥
⎣ .. . . 0 ⎦
0 · · · · · · 0 An

0 1
Ai = ,
−ωi2 0

C= C1 C2 · · · Cn ,

Ci = 1 0 ,
T T
x(t) = x1 (t) x2T (t) · · · xnT (t) ,

ai sin (ωi t + φi )
xi (t) = ,
ωi ai cos (ωi t + φi )
where x(t) is the state vector, y(t) is the measurement of q̇b (t), v(t) is the measure-
ment error, which is assumed to be a Gaussian white noise, i = 1, 2, . . . , n. This
model is time invariant, and is updated every ΔtFFT .
Then, this continuous-time model expressed by (7.2) and (7.3) is discretized as
follows:
xk+1 = Āxk , (7.4)

yk = C xk + vk , (7.5)
z k = C xk , (7.6)
where Ā = exp(AΔt), Δt is the sampling period, [·]k := [·](tk ), tk = kΔt, 0 ≤

k ∈ Z, z k = q̇bk to be estimated. Every ΔtFFT , the model is updated and k is reset
to 0. Note that this formulation does not include any system noise not as in the
conventional one.
7.2.2 Kalman Filtering Algorithm
The Kalman filtering algorithm for the system in (7.4)–(7.6) is given by
x̂k+1/k = Ā x̂k/k , (7.7)

f
x̂k/k = x̂k/k−1 + K k (yk − ẑ k/k−1 ), (7.8)
ẑ k+1/k = C x̂k+1/k , (7.9)
ẑ k/k = C x̂k/k , (7.10)
f
where xk is the state vector, K k is the Kalman gain, [·] ˆ i/j is the estimate of [·]i at the
f
jth time point, ẑ k/k is the kth output of the filter. The Kalman gain K k is obtained by
K k = Pk C T (C Pk C T + R)−1 ,
f
(7.11)
where R is the variance of vk , which is also updated every ΔtFFT , Pk is the covariance
matrix of the estimation error [xk − x̂k/k−1 ], which is recursively updated by the
recurrence formula:
f
Pk+1 = Ā(Pk − K k C Pk ) Ā T . (7.12)
The initial conditions x̂0/−1 , P0 and R are set as follows:
x̂0/−1 = x0 , (7.13)
1
0
P0 = x̃k x̃kT , (7.14)
n FFT
k=−n FFT +1
1
0
R= vk2 , (7.15)
n FFT
k=−n FFT +1
1
0
x̃k = xk − xj, (7.16)
n FFT
j=−n FFT +1

N FFT
vk = yk − ai sin (ωi tk + φi ) , (7.17)
i=1
where n FFT = TFFT /Δt set to be ∈ Z+ , xk , k = −n FFT + 1, −n FFT + 2, . . . , 0 is

obtained by the FFT-based model (7.4).
The advantage of the Kalman filter is to effectively reduce Gaussian white noise,
under the condition that the state-space model by Ā is available. However, if the
variation of the ship oscillation frequency is large and the current Ā is no longer
available, the estimation performance using the Kalman filter will deteriorate.
7.2.3 H∞ Filtering Algorithm
First, the suboptimal problem formulation of H∞ filter is expressed by

2
k z k − ẑ k
sup T < γ2. (7.18)
x0,v x0 − x̂0 P0−1 x0 − x̂0 + R −1 k vk
2
In our approach, γ is utilized as a filter design parameter, i.e., preparing multiple

H∞ filters with different γ’s. Note that the initial covariance matrix of estimation
error P0 and the covariance matrix of measurement error R are used as the weighting
function matrices in the above criterion in 7.18 instead.
Then, the H∞ filtering algorithm for the problem based on the prescribed system
is similar to that of the Kalman filter. The only difference is that the Kalman gain is
f
replaced by the H∞ filter gain Hk :
Hk = Pk C T (C Pk C T + R)−1 ,
f
(7.19)
Pk+1 = Ā Pk Ψk−1 Ā T , (7.20)
−1 −2
Ψk = I + (R −γ )C C Pk ,
T
(7.21)
where I is the identity matrix, γ is the design parameter. In particular, if γ → ∞,

the H∞ filtering algorithm coincides with the Kalman filtering one. Note that it is
in general difficult to appropriately set γ. Hence, we employ multiple filters with
different γ’s, which is the key point in the proposed algorithm. The number of the
employed H∞ filters is denoted by N .
The existence condition for the H∞ filter is that Pk > 0 and Ψk Pk−1 > 0 (k =
0, 1, 2, . . .). Further, one can easily confirm that R −1 − γ −2 ≥ 0 and P0 > 0 can be
its sufficient condition. Therefore, we define βγ := (R −1 −γ −2 )R (0 ≤ βγ ≤ 1), and
use βγ as the design parameter instead of γ in the proposed method. In this problem,
R is a scaler. βγ ’s are set as (m/N )2 (m = 0, 1, . . . , N ). In particular, when m = N ,
the filter is the Kalman filter.
The advantage of the H∞ filter is to effectively reduce bounded-norm determin-
istic noises and therefore is expected to be robust against the model-error, i.e., the
variation of ship oscillation frequency.
7.2.4 Method of Selectively Combining Multiple H∞

and Kalman Filters
To overcome the problem of model deterioration due to the ship oscillation frequency
change, we have proposed a method of selectively combining multiple H∞ and
Kalman filters, which aims at utilizing the advantage of each filter as much as possible.
The selection of filters are based on a criterion related to innovations, i.e., gaps
between measurements and estimates. For each filter, the Eqs. (7.8)–(7.10) can be
rewritten as follows:
ẑ k/k = αk yk + (1 − αk )ẑ k/k−1 , (7.22)

−1
αk = C Pk C T C Pk C T + R , (7.23)
which displays the role of the defined criterion denoted by α. In this form, α is
interpreted as a weight of contribution to the final estimate of the measurement yk
compared with that of ẑ k/k−1 . Thus, the smaller α implies that ẑ k/k−1 based on
the current model is the more reliable. Importantly, α will never reflect the actual
innovations but has been scheduled in advance once the filter has been designed.
Therefore, we have focused on this point and considered that if the predetermined
α is “in a sense” in accordance with the actual innovations then the current model
is appropriate and available. Note that C Pk C T is the variance of estimation error of
ẑ k/k−1 and C Pk C T + R is the variance of the innovation νk = yk − ẑ k/k−1 . Hence,
αk can be approximated by
−1
αk : = νi2 − R νi2 , (7.24)
1
k
νk2 := νi2 , (7.25)
nm
i=k−n m +1
where n m is the rounded integer of 2π/(ωmax Δt), ωmax is the angular frequency
with the maximum amplitude in (7.1).
As mentioned above, αk is predefined and does not contain any feedback of the
measurements once the model has been updated, whereas αk reflects the recent mea-
surements in the form of innovations, which includes the information on the model
variation between the model updates. Therefore, αk ≥ αk can be interpreted that the
filter actually works as expected or better than expected. Thus, using this condition
as an criterion, we select the filters at each time step and combine the respective
estimates by averaging them, which is central idea of our proposed methodology.
7.2.5 Integral with Drift Error Compensation
To acquire inclination angles, estimates of angular velocities via the algorithm are
numerically integrated by the trapezoidal rule. q̂b denotes the integral. The integral
is in general influenced by drift errors of the gyro sensor. Hence, the following drift
error compensation algorithm is used:

q̂bj ( j < TFFT /Δt + n m − 1)
q̂bj = , (7.26)
q̂bj − q̂b j ( j ≥ TFFT /Δt + n m − 1)
1
j

q̂b j = q̂bm , (7.27)
nm
m= j−n m +1
where n m is utilized as in the process of filter selection as prescribed before.

7.3 Ship Oscillation Motions
This section addresses ship oscillation motion for demonstration by simulation,

which is subject to variation of ship velocity. Specifically, an ocean wave spectrum
model and ship dynamical models are presented.
7.3.1 Ocean Wave Spectrum Model
As an ocean wave spectrum model, the Bretschneider spectrum [22] is employed with
a slight modification. See Sect. 2.4.2 again for the spectrum. In the demonstrations, we
use the wave spectrum model not as motion but forces driving the ship dynamics. By
linearly shifting ωi ’s as G w ωi ’s (G w is a constant multiplier) and linearly magnifying
the amplitudes Aωi ’s as G A Aωi (G A is a constant gain), two types of wave models
denoted by Model 1 and Model 2 are considered as shown in Table 7.1. Let f w denote
the normalized wave force expressed by
nω
f w (t) = Σi=1 G A Aωi sin(G w ωi t + φi ). (7.28)
Additionally, variation of ocean wave frequency due to ship velocity variation is

considered. Figure 7.2 depicts the variation rate of the frequency during the whole
simulation period of 160 s, which is divided into five periods and labeled with Periods
1–5, respectively. Figure 7.3 displays the respective wave force profiles.
7.3.2 Ship Dynamical Model
As an illustrative example of an equation of ship motion [69, 108], that of ship-rolling

motion is considered as in the following:
q̈b (t) + 0.50q̇b (t) + 5.0qb (t) + cqb3 (t) = f w (t) (7.29)
where qb (t) is the inclination angle, c is the constant coefficient representing the
nonlinearity, f w (t) is the normalized wave force mentioned above. When c = 0,
the natural damped frequency of this model is 0.71π rad/s, which corresponds to
Table 7.1 Ocean wave models

Model 1 Model 2
Peak frequency (rad/s) 0.72π 0.36π
Gw 3 1.5
GA 3 2
7.3 Ship Oscillation Motions 131
2.5
2
variation rate
1.5
Period 1 2 3 4 5
0.5
0 20 40 60 80 100 120 140 160
time (s)
Fig. 7.2 Variation process of the ocean wave frequency
4
accel. (rad/s2)
−2
−4
0 20 40 60 80 100 120 140 160
1
accel. (rad/s2)
0.5
−0.5
−1
0 20 40 60 80 100 120 140 160
time (s)
Fig. 7.3 Normalized external forces (acceleration); top Model 1, bottom Model 2
Table 7.2 Nonlinear parameter c of ship dynamical model

Nonlinear parameter Model A Model B Model C
c 0 200 −30
the peak frequency as shown in Table 7.1. This natural frequency is equal to the
peak frequency in Periods 1 and 5 with Model 1 and in Period 3 with Model 2. The
constant coefficient c is related to the ship model nonlinearity, namely, with c = 0
(7.29) is linear, with c > 0 (7.29) has a nonlinear stiffness of the hard-spring type,
and c < 0 provides a soft-spring type nonlinearity. According to the three cases of
c, three dynamical models named Model A, B and C are considered respectively as
presented in Table 7.2.
All the combinations of the three ship dynamical models and two wave models
will be demonstrated in simulations, which are referred to as Cases A1, A2, B1, B2,
C1, and C2, corresponding to the model indices.
7.4 Simulations
7.4.1 Conditions
In this section, the six model cases, A1–C2, mentioned in Sect. 7.3 are demonstrated
by simulations to evaluate the proposed estimation method. First, being based on the
respective wave and ship dynamical models, we performed ship motion simulations
with the initial condition of qb (0) = 0 and q̇b (0) = 0 and with the whole period of 160
s. Further, to construct test measurements of angular velocity, Gaussian white noises
with the mean of 0◦ /s and the standard deviation of 8◦ /s were added to the obtained
angular velocity data as sensor noises. Then, these test measurements were supplied
to the proposed estimation system. The parameters set for the estimation algorithm
are shown in Table 7.3. TFFT and ΔtFFT approximately equal to four and eight times
the natural period of the ship oscillation, respectively. Under those conditions, we
have conducted simulations and compared the respective estimation performances
of the proposed method employing 10 H∞ and single Kalman filters, the integral
with drift error compensation with no filter, from the viewpoint of root mean square
error (RMSE) of estimation in the respective Periods 1–5, and in the whole period.
Table 7.3 Parameters for the estimation algorithm in simulations

Δt TFFT ΔtFFT n N
0.1 s 11.2 s 22.4 s 5 10
Δt sampling period, TFFT FFT period, ΔtFFT model update period, n number of modes, N number
of H∞ filters
7.4 Simulations 133
10
angle (deg)
0
−10
0 20 40 60 80 100 120 140 160
10
angle (deg)
−10
0 20 40 60 80 100 120 140 160
10
angle (deg)
−10
0 20 40 60 80 100 120 140 160
time (s)
Fig. 7.4 Ship inclination angles; top Case A1, second from the top B1, bottom C1
7.4.2 Results
First let us see the simulation results of ship motion in the respective cases as shown
in Figs. 7.4 and 7.5. Figure 7.4 compares the three ship dynamical models with the
wave model Model 1. The features of the profiles are similar despite the differences
in the dynamical models, where at t = 50−110 (Period 2–Period 4) the amplitude
has dramatically decayed. Comparing the dynamical models, in Periods 1 and 5 the
amplitude of Case B1 (hard-spring) is the largest, that of A1 (linear) is the second,
and that of C1 (soft-spring) is the smallest. On the other hand, in Fig. 7.5 (for Model
2), quite different profiles from those in the case of Model 1 can be observed, where
the amplitudes do not vary much according to the periods. These results are due to
the property of low-pass filter of the ship dynamical models. Namely, waves with
frequencies much greater than the natural damped frequency near 0.71π rad/s is
difficult to be passed whereas waves with frequencies less than that can be passed
easily. In terms of comparison of the models, the tendency is the opposite from that
for Model 1. In Periods 1 and 5, the amplitude of Case C2 (soft-spring) is the largest,
that of A2 (linear) is the second, and that of B2 (hard-spring) is the smallest.
Next, we examine the simulation results in the form of root mean square errors
(RMSEs) of estimation which are shown in Figs. 7.6, 7.7, 7.8, 7.9, 7.10 and 7.11
for the respective model cases. In each figure, the circle represents the RMSE of
the proposed method, * is for that of the best single filter, + is for that of the worst
single filter, the triangle is for that of the integral with drift compensation without a
10
angle (deg)
0
−10
0 20 40 60 80 100 120 140 160
10
angle (deg)
−10
0 20 40 60 80 100 120 140 160
10
angle (deg)
−10
0 20 40 60 80 100 120 140 160
time (s)
Fig. 7.5 Ship inclination angles; top Case A2, second from the top B2, bottom C2
Fig. 7.6 Simulation results 5

of RMSE for Case A1 (linear
ship model); Period 6 4
represents all the periods
RMSE (deg)
1–5,
: proposed filter, *: 3
the best filter, +: the worst
filter, : integral 2
compensated
1
0
1 2 3 4 5 6
Period

of RMSE for Case B1
(nonlinear ship model); 4
Period 6 represents all the
RMSE (deg)
periods 1–5,
: proposed 3
filter, *: the best filter, +: the
worst filter, : integral 2
compensated
1
0
1 2 3 4 5 6
Period
7.4 Simulations 135

of RMSE for Case C1
RMSE (deg)
periods 1–5,
: proposed 3
compensated
1
0
1 2 3 4 5 6
Period

of RMSE for Case A2 (linear
ship model); Period 6 4
represents all the periods RMSE (deg)
1–5,
: proposed filter, *: 3
the best filter, +: the worst
filter, : integral 2
compensated
1
0
1 2 3 4 5 6
Period

of RMSE for Case B2
RMSE (deg)
periods 1–5,
: proposed 3
compensated
1
0
1 2 3 4 5 6
Period
filter. Period 6 means the whole period (Periods 1–5). Let us see Figs. 7.6, 7.7 and
7.8 which display the results in the case of Model 1 and compare the corresponding
ship oscillation data in Fig. 7.4. First, by focusing the best single filter (*), compare
the RMSEs and the corresponding ship oscillation carefully. Then, it is seen that the
RMSEs are proportional to the amplitudes of ship oscillation, and are not influenced
by the situations of time-varying and time-invariant oscillation frequency, nor by
the nonlinearities in the ship dynamical models. Whereas the other RMSEs exhibit
similar tendencies but which are not as clear as that of the best single filter. It should be

of RMSE for Case C2
RMSE (deg)
periods 1–5,
: proposed 3
compensated
1
0
1 2 3 4 5 6
Period
noted that the best filter depends on situations and cannot be determined in advance.
On the other hand, the performances of the worst filter is quite poor, which implies
how influential the filter design parameter βγ will be to the resulting estimation
performance. Then, let us investigate the performances of the proposed method.
From the figures, it is observed that in most cases the proposed method exhibit
similar good performances to those of the best filter, except in Period 2, which has
proven that the filter selection method base on the criterion α is effective. Comparing
the proposed method with the integral with drift error compensation, the proposed
method reveals superior performance.
Moreover, we examine Figs. 7.9, 7.10 and 7.11 for the case of Model 2. Similarly,
we compare the RMSEs in those figures and the corresponding ship oscillation data
in Fig. 7.5. Then, we notice the same argument can be applied to this case of Model
2 as in the case of Model 1. The performance of the best filter is only influenced by
the signal amplitude, and the proposed method can gives very near performance to
that of the best filter and is superior to the integral with drift error compensation with
no filter.
Consequently, from all the above arguments, we have confirmed that the proposed
estimation algorithm performs successfully in the presence of oscillation frequency
variation and even nonlinearity in the ship dynamics by effectively selecting appro-
priate filters from the multiple filters.
7.5 Conclusions
In this chapter, an estimation method of ship oscillation in the presence of oscilla-

tion frequency variation has been presented, assuming that only an angular velocity
is measured with a low-cost rate gyro sensor. This method incorporates periodic
updates of the FFT-based model and the method of selectively combining multiple
H∞ and Kalman filters according to the innovation-based criterion. To evaluate the
estimation performance of the proposed method, simulations have been conducted
with consideration of the variation of ship oscillation frequency and nonlinearities
7.5 Conclusions 137
in the ship dynamics. The results have proven that the proposed method is effective
and robust under such difficult conditions, and justified the idea of combining filters
aiming at utilizing the advantage of each filter. However, there still remains a room
for improvement with respect to the filter selection method, which the results in
Period 2 in the simulations have suggested. In fact, we are planning adding another
type of criterion to the current innovation-based one.
References
1. Abdallah C, Dawson D, Dorato P, Jamshidi M (1991) Survey of robust control for rigid robots.
IEEE Control Syst Mag 11:24–30
2. Alves RM, Battista RC (1999) Active control of heave motion for tlp type offshore platform.
In: Proceedings of the international offshore and polar engineering conference, Brest, France,
pp 332–338
3. Apkarian P, Chretien JP, Gahinet P, Biannic JM (1993) μ synthesis by D-K iterations with
constant scaling. In: Proceedings of the american control conference, pp 3192–3196
4. Balas GJ, Doyle JC, Glover K, Packard A, Smith R (1995) μ-Analysis and synthesis toolbox
user’s guide. MUSYN Inc. and The Math Works Inc, Minneapolis
5. Bandyopadhyay B, Deepak F, Kim K-S (2009) Sliding mode control using novel sliding
surfaces. LNCIS, vol 392. Springer, Heidelberg
6. Batista P, Silvestre C, Oliveira P (2008) Kalman and H∞ optimal filtering for a class of
kinematic systems. In: Proceedings of the 17th world congress in IFAC, pp 12528–12533
7. Battista RC (1999) Active control of heave motion for tlp type offshore platform under random
waves. Smart Structures and Materials 1999: Smart Systems for Bridges. Structures and
Highways, Newport Beach, USA, pp 184–193
8. Becker N, Grimm WM (1988) On L 2 - and L ∞ -stability approaches for the robust control of
robot manipulators. IEEE Trans Autom Control 33:118–122
9. Bowden KF (1983) Physical oceanography of costal waters. Ellis Horwood Limited, Chich-
ester
10. Chida Y, Yamaguchi Y, Soga H, Kida T, Yamaguchi I, Sekiguchi T (1996) On-orbit attitude
control experiments for ETS-VI–I-PD and two-degree-of-freedom H∞ control–. In: Proceed-
ings of the IEEE conference on decision and control, pp 486–488
11. Choi SB, Park DW (1994) Moving sliding surfaces for fast tracking control of second-order
dynamic systems. ASME J Dyn Syst Meas Control 116:154–158
12. Choi SB, Park DW, Jayasuriya S (1994) A time-varying sliding surface for fast and tracking
control of second-order dynamic systems. Automatica 30:899–904
13. Cloutier JR, D’Souza CN, Mracek CP (1996) Nonlinear regulation and nonlinear H∞ control
via the state-dependent Riccati equation technique: part 1, theory. In: Proceedings of the
international conference on nonlinear problems in aviation and aerospace, pp 117–130
14. Desoer CA, Vidyasagar M (1975) Feedback systems: input-output properties. Academic, New
York
15. Do KD, Pan J (2008) Nonlinear control of an active heave compensation system. Ocean Eng
35:558–571

DOI 10.1007/978-3-319-21780-2
140 References
16. Doyle JC, Glover K, Khargonekar PP, Francis BA (1989) State-space solutions to standard
H∈ and H∞ control problems. IEEE Trans Autom Control 34:831–847
17. Doyle J, Packard A, Zhou K (1991) Review of LFTs, LMIs, and μ. In: Proceedings of the
IEEE conference on decision and control, pp 1227–1232
18. Dubowsky S, Papadopoulos E (1991) The kinematics, dynamics, and control of free-flying
and free-floating space robotic systems. IEEE Trans Robot Autom 9(5):531–543
19. Fadali MS, Yaz E (1995) Stability robustness and robustification of the exact linearization
method of robotic manipulator control. In: Proceedings of the IEEE conference on decision
and control, pp 1624–1629
20. Fallaha CJ, Saad M, Kanaan HY, Al-Haddad K (2011) Sliding-mode robot control with
exponential reaching law. IEEE Trans Ind Electron 58(2):600–610
21. Fang Y, Wang P, Sun N, Zhang Y (2014) Dynamics analysis and nonlinear control of an
offshore boom crane. IEEE Trans Ind Electron 61(1):414–427
22. Fossen TI (1994) Guidance and control of ocean vehicles. Wiley, Chichester
23. Fossen TI, Perez T (2009) Kalman filtering for positioning and heading control of ships and
offshore rigs. IEEE Control Syst Mag 32–46
24. Francsis B (1987) A course in H∞ control theory. Springer, New York
25. Furuta K (1990) Sliding mode control of a discrete system. Syst Control Lett 14:145–152
26. Furuta K (1993) VSS type self-tuning control. IEEE Trans Ind Electron 40(1):37–44
27. Gérard B, Souley H, Zasadzinski M, Darouach M (2011) H∞ filter for bilinear systems using
LPV approach. Sig Process 91:1168–1181
28. Glover K, Doyle JC (1988) State-sspace formulae for all stabilizing controllers that satisfy
and H∞ -norm bound and relations to risk sensitivity. Syst Control Lett 11:167–172
29. Godhavn JM (1998) Adapting tuning of heave filter in motion sensor. In: Proceedings of the
OCEANS, pp 174–178
30. Gu D-W, Petkov PH (2013) Robust control desing with MATLAB® , 2nd edn. Springer,
London
31. Gu K, Zohdy MA, Loh NK (1990) Necessary and sufficient conditions of quadratic stability
of uncertain linear systems. IEEE Trans Autom Control 35:601–604
32. Ha QP, Rye DC, Durrant-Whyte HF (1999) Fuzzy moving sliding mode control with appli-
cation to robotic manipulators. Automatica 35(4):607–616
33. Hassibi B, Sayed AH, Kailath T (1996) Linear estimation in Krein spaces–part II: applications.
IEEE Trans Autom Control 41(1):34–49
34. El-Hawary F, Mbamalu GAN (1996) Dynamic heave compensation using robust estimation
techniques. Comput Electr Eng 22(4):257–273
35. Hootsmans NAM, Dubowsky S (1991) Large motion control of mobile manipulators including
vehicle suspension characteristics. In: Proceedings of the IEEE international conference on
robotics and automation, Sacramento, USA, pp 2336–2341
36. Hootsmans NAM, Dubowsky S, Mo PZ (1992) The experimental performance of a mobile ma-
nipulator control algorithm. In: Proceedings of the IEEE international conference on robotics
and automation, pp 1948–1954
37. Hosoe S, Zhang T (1988) An elementary state space approach to R H ∞ optimal control. Syst
Control Lett 11:369–380
38. Huang S, Vassalos D (1995) Optimal control of the heave motion of marine cable subsea-unit
systems. In: Proceedings of the international offshore and polar engineering conference, The
Hague, Netherlands, pp 203–208
39. Isram S, Liu XP (2011) Robust sliding mode control for robot manipulators. IEEE Trans Ind
Electron 58(6):2444–2453
40. Isidori A (1995) Nonlinear control systems. Springer, Berlin
41. Itkis U (1976) Control systems of variable structure. Wiley, New York
42. Iwamura T, Toda M (2013) Motion control of an oscillatory-base manipulator using sliding
mode control via rotating sliding surface with variable-gain integral control. In: Proceedings
of the american control conference, Washington DC, USA, pp 5762–5767
References 141
43. Iwasaki T (1994) Robust performance analysis for systems with norm-bounded time-varying
structured uncertainty. In: Proceedings of the american control conference, Maryland, pp
2343–2347
44. Iwasaki T, Skelton RE (1994) All controllers for general H∞ control problem: LMI existence
conditions and state space formulas. Automatica 30(8):1307–1317
45. Johansen TA, Fossen TI, Sagatun SI, Nielsen FG (2003) Wave synchronizing crane control
during water entry in offshore moonpool operations-experimental results. IEEE J Ocean Eng
28(4):720–728
46. Joshi J, Desrochers AA (1986) Modeling and control of a mobile robot subject to disturbances.
In: Proceedings of the IEEE international conference on robotics and automation, pp 1508–
1513
47. Karimi HR (2009) Robust H∞ filter design for uncertain linear systems over network with
network-induced delays and output quantization. Model Identif Control 30(1):27–37
48. Khalil HK (2002) Nonlinear systems, 3rd edn. Prentice Hall, Upper Saddle River
49. Kharrati H, Khanmohammadi S (2008) Genetic algorithm combined with H∞ filtering for
optimizing fuzzy rules and membership function. J Appl Sci 8(19):3439–3445
50. Kimura H (1981) Robust stabilizability for a class of transfer functions. IEEE Trans Autom
Control 27:783–793
51. Kimura H (1989) Conjugation, interpolation and model-matching in H∞ . Int J Control
49:269–307
52. Korde UA (1998) Active heave compensation on drill-ships in irregular waves. Ocean Eng
25(7):541–561
53. Kreutz K (1989) On manipulator control by exact linearization. IEEE Trans Autom Control
34:763–767
54. Küchler S, Mahl T, Neupert J, Schneider K, Sawodny O (2011) Active control for an offshore
crane using prediction of the vessel’s motion. IEEE-ASME Trans Mechatron 17(2):297–309
55. Leavitt J, Sideris A, Bobrow JE (2011) Accurate tilt sensing with linear model. IEEE Sens J
11(10):2301–2309
56. Levant A (1993) Sliding order and sliding accuracy in sliding model control. Int J Control
58(6):1247–1263
57. Lew JY, Moon SM (1999) Acceleration feedback control of compliant base manipulators. In:
Proceedings of the american control conference, pp 1955–1958
58. Lew JY, Moon SM (2001) A simple active damping control for compliant base manipulators.
IEEE/ASME Trans Mechatron 6:305–310
59. Li W, Jia Y (2010) H-infinity filtering for a class nonlinear discrete-time systems based on
unscented transform. Signal Process 90:3301–3307
60. Liang Y-W, Xu S-D, Liaw D-C, Chen C-C (2008) A study of T-S model-based SMC scheme
with application to robot control. IEEE Trans Ind Electron 55(11):3964–3971
61. Lin J, Huang ZZ (2007) A hierarchical fuzzy approach to supervisory control of robot ma-
nipulators with oscillatory bases. Mechatoronics 17:589–600
62. Lin J, Lin CC, Lo H-S (2009) Psedo-inverse jacobian control with grey relational analysis for
robot manipulators mounted on oscillatory bases. J Sound Vib 326:421–437
63. Liu Z, Yang Q, Guo Z, Li J (2011) An improved autoregressive method with Kalman filtering
theory for vessel motion prediction. Intell Eng Syst 4(4):11–18
64. Jimenez-Lozano N, Goodwine B (2010) Nonlinear disturbance decoupling for a nonholo-
nomic mobile robotic manipulation platform. In: Proceedings of the IEEE international con-
ference on control. automation, robotics and vision, pp 1530–1535
65. Magee DP, Book WJ (1995) Filtering micro-manipulator wrist commands to prevent flexible
base motion. In: Proceedings of the american control conference, pp 924–928
66. Mahony R, Hamel T, Pflimlin J-M (2008) Nonlinear complementary filters on the special
orthogonal group. IEEE Trans Autom Control 53(5):1203–1218
67. Marconi L, Isidori A, Serrani A (2002) Autonomous vertical landing on an oscillating plat-
form: an internal-model based approach. Automatica 38:21–32
142 References
68. Megretski A (1993) Necessary and sufficient conditions of stability: a multi-loop generaliza-
tion of the circle criterion. IEEE Trans Autom Control 38:753–756
69. Morrall A (1980) The Gaul disaster: an investigation into the loss of a large stern trawler.
Trans R Inst Nav Archit 123:391–416
70. Nenchev DN, Yoshida K, Vichitkulsawat P, Konno A, Uchiyama M (1997) Experiments
on reaction null-space based decoupled control of a flexible structure mounted manipulator
system. In: Proceedings of the IEEE international conference on robotics and automation, pp
2528–2534
71. Neupert J, Mahl T, Haessig B, Sawodny O, Schneider K (2008) A heave compensation ap-
proach for offshore cranes. In: Proceedings of the american control conference, Seattle, USA,
pp 538–543
72. Packard A, Doyle J (1993) The complex structured singular value. Automatica 29:71–109
73. Ngo QH, Hong K-S (2012) Sliding-mode antisway control of an offshore container crane.
IEEE-ASME Trans Mechatron 17(2):201–209
74. Papadopoulos E, Dubouwsky S (1991) On the nature of control algorithms for free-floating
space manipulators. IEEE Trans Robot Autom 7(6):750–758
75. Pernebo L (1981) An algebraic theory for the design of controllers for linear multivari-
able systems–part I: structure matrices and feedforward design. IEEE Trans Autom Control
26:171–182
76. Poolla K, Tikku A (1995) Robust performance against time-varying structured perturbations.
IEEE Trans Autom Control 40(9):1589–1602
77. Qu Z, Dawson DM (1996) Robust tracking control of robot manipulators. IEEE Press, New
Jersey
78. Rho H, Kang Y, Hyun S, Kim H (2010) Discrete H∞ estimator design of unknown input:
game-theoretic approach. IEEE Trans Autom Control 55(7):1674–1688
79. Rotea MA, Iwasaki T (1994) An alternative to the D − K iteration?. In: Proceedings of the
american control conference, pp 53–57
80. Sampei M, Mita T, Nakamichi M (1990) An algebraic approach to H∞ output feedback
control problems. Syst Control Lett 14:13–24
81. Sanca A, Ferreira JP, Javier P (2012) A real-time attitude estimation scheme for hexarotor
micro aerial vehicle. ABCM Symp Ser Mechatron 5:1160–1166
82. Sato M, Toda M (2009) Motion control of an oscillatory-base manipulator in the global
coordinates. In: proceedings of the IEEE international conference control and automation,
New Zealand, pp 349–354
83. Sato M, Toda M (2015) Robust motion control of an oscillatory-base manipulator in a global
coordinate system. IEEE Trans Ind Electron 62(2):1163–1174
84. Sato M, Toda M (2014) Estimation of ship oscillation subject to ship speed variation using an
algorithm combining H∞ and Kalman Filters. In: Proceedings IEEE international conference
control and automation, Taichung, Taiwan, pp 929–934 (IEEE Trans Ind Electron 62(2):1163–
1174 (2015))
85. Shaked U, Theodor Y (1992) H∞ -optimal estimation: a tutorial. In: Proceedings of the IEEE
international conference decision and control. Tucson, Arizona, pp 2278–2286
86. Shamma JS (1994) Robust stability with time-varying structured uncertainty. IEEE Trans
Autom Control 39:714–724
87. Sharon A, Hardt D (1984) Enhancement of robot accuracy using end-point feedback and a
macro-micro manipulator system. In: Proceedings of the american control conference, pp
1836–1842
88. Skaare B, Egeland O (2006) Parallel force/position crane control in marine operations. IEEE
J Ocean Eng 31(3):599–613
89. Spong MW (1992) On the robust control of robot manipulators. IEEE Trans Autom Control
37:1782–1786
90. Spong MW, Vidyasagar M (1987) Robust linear compensator design for nonlinear robotic
control. IEEE J Robot Autom 3:345–351
91. Spong MW, Vidyasagar M (1989) Robot dynamics and control. Wiley, New York
References 143
92. Stein G, Doyle JC (1991) Beyond singular values and loop shapes. J Guid 14:5–16
93. Tikku A, Poolla K (1993) Robust performance against slowly-varying structured perturba-
tions. In: Proceedings of the IEEE conference on decision and control, pp 990–995
94. Toda M (1999) Robust control for mechanical systems with oscillating bases. In: Proceedings
of the IEEE international conference on systems man cybernetics, vol 2. Tokyo, Japan, pp
878–883
95. Toda M (2004) An H∞ control-based approach to robust control of mechanical systems with
oscillatory bases. IEEE Trans Robot Autom 20(2):283–296
96. Toda M (2004) A unified approach to control of mechanical systems with a flexible structure.
In: Proceedings of the international symposium on robotics and automation, Mexico, pp 313–
319
97. Toda M (2007) A unified approach to robust control of flexible mechanical systems. In:
Proceedings of the IEEE international conference on decision and control, New Orleans,
USA, pp 5787–5792
98. Toda M (2010) A unified approach to robust control of flexible mechanical systems using
H∞ control powered by PD control. In: Shafiei SE (ed) Advanced Strategies for Robot
Manipulators, Sciyo, Rijeka, Croatia, ch 13. pp 273–286
99. Torres MA, Dubowsky S, Pisoni AC (1994) Path-planning for elastically-mounted space
manipulators: experimental evaluation of the coupling map. In: Proceedings of the IEEE
international conference on robotics and automation, pp 2227–2233
100. Triantafyllou MS, Bodson M, Athans M (1983) Real time estimation of ship motions using
Kalman filtering techniques. IEEE J Eng OE-8(1):9–20
101. Utkin VI (1977) Variable structure system with sliding mode. IEEE Trans Autom Control
22(2):212–221
102. Utkin VI (1992) Sliding mode in control optimization. Springer, New York
103. Utkin VI (1993) Sliding mode control design principles and applications to electric drives.
IEEE Trans Ind Electron 40(1):23–36
104. Vidyasagar M, Kimura K (1986) Robust controllers for uncertain linear multivariable systems.
Automatica 22:85–94
105. Wakasa Y, Yamamoto Y (1999) Control system design considering a tradeoff between eval-
uated uncertainty ranges and control performance. Asian J Control 1:49–57
106. Wang QC, Li J, Zhang MX, Yang CH (2011) H-infinity filter based particle filter for maneu-
vering target tracking. Prog Electromagn Res B 30:103–116
107. Wang XF, Zhao XR, Yang XJ (2006) Research of motion resolving and filtering algorithm of
a ship’s three-freedom motion simulation platform based on LabVIEW. Instrum Sci Technol
48:149–153
108. Wright JHG, Marshfield WB (1979) Ship roll response and capsize behavior in beam seas.
Trans R Inst Nav Archit 122:129–144
109. Yamada Y, Hara S (1998) Global optimization for H∞ control with constant diagonal scaling.
IEEE Trans Autom Control 43:191–203
110. Yamamoto Y, Yun X (1994) Coordinating locomotion and manipulation of a mobile manip-
ulator. IEEE Trans Autom Control 39(6):1326–1332
111. Yang J, Li S, Yu X (2013) Sliding-mode control for systems with mismatched uncertainties
via a disturbance observer. IEEE Trans Ind Electron 60(1):160–169
112. Yun X, Aparicio C, Bachmann ER, McGhee RB (2005) Implementation and experimental
results of a quaternion-based kalman filter for human body motion tracking. In Proceedings
of the IEEE international conference on robotics and automation, Barcelona, pp 317–322
113. Zames G (1981) Feedback and optimal sensitivity: Model reference transformations, multi-
plicative seminorms, and approximate inverses. IEEE Trans Autom Control 26:301–320
114. Zames G, Francis BA (1983) Feedback, minimax sensitivity, and optimal robustness. IEEE
Trans Autom Control 28:585–601
115. Zhou K, Doyle JC, Glover K (1995) Robust and optimal control. Prentice Hall, New Jersey
Index
Symbols D
D–K iteration, 27, 47, 51 Drift error compensation, 129
with constant scalings, 27
γ -iteration, 25, 47, 51
μ with constant scalings, 60 E
μ-analysis, 26 Estimation algorithm, 124
μ-synthesis, 27 Experimental OBM, 15
H∞ control, 22, 42, 83, 88, 108, 109, 118 Extended matrix polytope, 28, 30, 32, 36
mixed-sensitivity problem, 44
optimal control problem, 24
standard problem, 24, 42
suboptimal control problem, 24 F
synthesis, 51 FFT, 124
H∞ filter, 124, 127 Frobenius norm, 44
H∞ filter gain, 128
H∞ norm, 22
1 :criterion for the hyperplane, 40 G
2 :criterion for the hyperplane, 40 Gaussian white noise, 127
vec(·), 39 Generalized plant, 23, 25, 44
vec−1 (·), 39 system configuration, 48
H2 norm, 22 Global-coordinate problems, 9
L2 gain, 22 Gyro sensor, 129
A H
Approximation error vector, 40 Heave motion, 8
Attitude control, 9, 69
I
B Illustrative model, 7, 36, 98, 99
Base oscillation model, 12, 68 Influence of sensor error, 80
Bretschneider oscillation, 13 Innovation, 128
double-frequency oscillation, 12 Internal model principle, 2
single-frequency oscillation, 12 Inverse kinematics, 70
Bretschneider spectrum, 12, 130 Itkis, 97
DOI 10.1007/978-3-319-21780-2
146 Index
K Reduced-order controller, 53
Kalman filter, 124, 126 Region of attraction, 64
Kalman gain, 127 Robust control toolbox, 25
Robustness analysis, 55
Robust performance, 88, 118
L Robust stability, 26
Lagrangian mechanics, 11 for structured uncertainties, 26
LFT, 23 for unstructured uncertainties, 25
lower LFT, 24 Robust stability analysis, 62
upper LFT, 23 Root mean square error, xi, 70, 132
Linear fractional transformation, xi Rotating sliding surface with variable-gain
Linear quadratic, xi integral control, xi, 99
Linear state-feedback control, 42 RSSI
Local-coordinate problems, 9 global stability, 105
LQ control, 22 semi-global stability, 106
Lyapunov function, 62, 64, 103 stability analysis, 102
M S
Manipulator dynamics, 10 Sampling period, 17
Matrix polytope, 29 Scaling matrices, 52
Model reduction, 53 Selectively combining method, 128
Multi-input multi-output, xi Semi-global internal stability, 62
Sensitivity functions, 41, 55
complimentary sensitivity function, 55
N quasi-complementary sensitivity func-
Nominal-case performance, 68, 109 tion, 42
Nominal parameter, 19 sensitivity function, 42
Nonlinear state-feedback control, 34 Sensor resolution, 17, 80
Notch-type filter, 46 Ship oscillation model, 130
Single-input single-output, xi
Sliding condition, 99
O Sliding-mode control, 97
Offshore mechanical systems, 2, 28 SMC-I, 99, 101
Orthogonal projection, 39 SMC-RSSI, 99, 101, 108, 109, 118
Oscillatory-base manipulator (OBM), 1 Small-gain theorem, 23
Stable
asymptotically stable, 64
P internally stable, 23
Parameter variation, 17 robustly stable, 26
Payload variation, 17, 28 State-dependent coefficient form, 55, 62
Payload width, 88 Structured singular value, 26
PID control, 69, 83, 88 Structured uncertainty, 26
Poincare-Bendixon theorem, 107 System poles, 53
Position control, 9, 70 System zeros, 53
Power spectral density, xi, 80
T
Q Time-invariant uncertainty, 41
Quadratic stability, 65 Time-varying uncertainty, 41
R U
Redheffer star product, 30, 31 Uncertainty set, 50
Index 147
Unstructured uncertainty, 26 W
Utkin, 97 Weighting function, 44
V
Variable structure system, xi, 97 Z
Virtual linear plant, 34, 42 Zames, 22

Robust Motion Control of Oscillatory-Base Manipullators

Uploaded by

Copyright:

Available Formats

Robust Motion Control of Oscillatory-Base Manipullators

Uploaded by

Document Information

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

Robust Motion Control of Oscillatory-Base Manipullators

Uploaded by

Copyright:

Available Formats

Lecture Notes in Control and Information Sciences 463

Series Advisory Boards

More information about this series at http://www.springer.com/series/642

Robust Motion Control

Additional material to this book can be downloaded from http://extras.springer.com.

ISSN 0170-8643 ISSN 1610-7411 (electronic)

Library of Congress Control Number: 2015944735

Mathematics Subject Classiﬁcation: 70Q05

Springer Cham Heidelberg New York Dordrecht London

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

The subject of the monograph is robust control of “An oscillatory-base manipulator

Please note that the monograph has been designed to be as self-contained as

Tokyo, Japan Masayoshi Toda

2 Problem Definition, Dynamical Model Formulation . . . . . . . . . . . . 7

3 Experimental Apparatus and Analysis on Parameter Variation

4 Motion Control Using an H1 -Control-Based Approach . . . . . .... 21

4.4 Control Design and Analysis . . . . . . . . . . . . . . . . . . . . . ..... 33

5 Simulations and Experiments for the H1 -Control-Based

6 Motion Control Using a Sliding-Mode-Control-Based Approach . .. 97

7 Base Oscillation Estimation via Multiple H1 Filters . . . . . . . . . . . 123

1.1 What Is an Oscillatory-Base Manipulator?

“An oscillatory-base manipulator (OBM)” which we will focus on throughout this

© Springer International Publishing Switzerland 2016 1

environments are considerably difficult for mechanical systems because of wave,

1.2 Previous Works and Contents of the Monograph

• Chap. 2 presents the problem definition, including important assumptions, the

2.2 Problem Definition

As an illustrative model of OBMs, we consider a two-DOF manipulator with a one-

Fig. 2.1 Schematic diagram Y

employed is because, in practical cases of offshore mechanical systems, roll and

2.3 Local-Coordinate and Global-Coordinate Problems

Fig. 2.2 Position control in a base-fixed coordinate system (local case)

Fig. 2.3 Attitude control in a base-fixed coordinate system (local case)

Fig. 2.4 Position control in a global-coordinate system

Fig. 2.5 Attitude control in a global-coordinate system

2.4 Dynamical Model Formulation

2.4.1 Dynamical Model of an OBM

The dynamics of n-rigid link manipulators with revolutionary joints subject to

M(q)q̈ + C(q, q̇)q̇ + D q̇ + G(q, q b ) + H (q, q̇, q̇ b , q̈ b ) = τ , (2.1)

where q ∈ Rn and q b ∈ Rm denote the position vectors of the manipulator links

H (q, q̇, q̇ b , q̈ b ) = [H1 , H2 ]T

As in (2.7), the base oscillation is modeled as a linear combination of multiple

2.4.2 Base Oscillation Model

Fig. 2.6 Power spectrum of the Bretschneider model

Abstract We developed an experimental apparatus in order to evaluate control sys-

This chapter introduces an apparatus that we developed for experimental evaluation

3.2 Experimental OBM

In order to perform experimental evaluations on control systems for OBMs, we have

Fig. 3.1 Experimental

Table 3.1 Physical parameters of the experimental OBM

Table 3.2 Specifications of the motors installed on the experimental OBM

control system analyses, and simulation demonstrations of control system perfor-

3.3 Analysis on Parameter Variation Due to Payloads

This section analyzes parameter variation in the dynamical model of manipulator

Fig. 3.2 Sample of payloads

In the inertia matrix M(q),

Hence, the centripetal and Coriolis term is rewritten as:

ΔM(q) = M(q) − Mn (q)

Table 3.4 Parameter variations due to change of the payload

where X a1 = diag[β1 ,β2 ], X a2 = diag[−β1 , β2 ], X a3 = −X a1 , X a4 = −X a2 , λi

where X b1 = diag[β1 , 0], X b2 = X a1 , X b3 = diag[0, β2 ], νi ∈ R, 0 ≤ νi ≤ 1 for

where X c1 = diag[β1 , β2 ], X c2 = diag[β1 , β2 ], κi ∈ R, 0 ≤ κi ≤ 1 for i = 1, 2,

wherel is the number of the constant vertex matrices X i ’s, 0 ≤ λi ≤ 1 (i = 1, . . . , l)

ΔX = Fu (S(Ψ2 , Ψ1 ), Δ), (4.27)

where · denotes the Euclidean norm. 1 indicates how close to a hyperplane

2 = max σ̄ (vec−1 (d I − d I ∗ )), (4.50)