NEURAL NETWORKS
IN ROBOTICS
T H E K L U W E R INTERNATIONAL SERIES
IN ENGINEERING AND COMPUTER SCIENCE
ROBOTICS: VISION, MANIPULATION AND SENSORS
Consulting Editor: Takeo Kanade
VISION AND NAVIGATION: THE CMU NAVLAB, C. Thorpe (editor)
ISBN 0-7923-9068-7
TASK-DIRECTED SENSOR FUSION AND PLANNING: A Computational Approach,
G. D. Hager
ISBN: 0-7923-9108-X
COMPUTER ANAL YSIS OF VISUAL TEXTURES, F. Tomita and S. Tsuj i
ISBN: 0-7923-9114-4
DATA FUSION FOR SENSORY INFORMATION PROCESSING S YSTEMS, J. Clark
and A. Yuille
ISBN: 0-7923-9120-9
PARALLEL ARCHITECTURES AND PARALLEL ALGORITHMS FOR INTEGRATED
VISION SYSTEMS, A . N . Choudhary, J. H. Patel
ISBN: 0-7923-9078-4
ROBOT MOTION PLANNING, J. Latombe
ISBN: 0-7923-9129-2
DYNAMIC ANALYSIS OF ROBOT MANIPULATORS: A Cartesian Tensor Approach,
CA Balafoutis, R.V. Patel
ISBN: 07923-9145-4
PERTURBATION TECHNIQUES FOR FLEXIBLE MANIPULATORS: A. Fraser and
R. W. Daniel
ISBN: 0-7923-9162-4
COMPUTER AIDED MECHANICAL ASSEMBLY PLANNING: L. Homen de Mello and
S. Lee
ISBN: 0-7923-9205-1
INTELLIGENT ROBOTIC SYSTEMS FOR SPACE EXPLORATION: Alan A. Desrochers
ISBN: 0-7923-9197-7
MEASUREMENT OF IMAGE VELOCITY: David J. Fleet
ISBN: 0-7923-9198-5
DIRECTED SONAR SENSING FOR MOBILE ROBOT NAVIGATION: John J. Leonard,
Hugh F. Durrant-Whyte
ISBN: 0-7923-9242-6
A GENERAL MODEL OF LEGGED LOCOMOTION ON NATURAL TERRAIN: David
J. Manko
ISBN: 0-7923-9247-7
INTELLIGENT ROBOTIC SYSTEMS: THEORY, DESIGN AND APPLICATIONS K.
Valavanis, G. Saridis
ISBN : 0-7923-9250-7
QUALITATIVEMOTION UNDERSTANDING: W. Burger, B. Bhanu
ISBN: 0-7923-9251-5
DIRECTED SONAR SENSING FOR MOBILE ROBOT NAVIGATION: J.J. Leonard,
H.F. Durrant-Whyte
ISBN: 0-7923-9242-6
NONHOLONOMIC MOTION PLANNING: Z. L i , J.F. Canny
ISBN: 0-7923-9275-2
SPACE ROBOTICS-.DYNAMICS AND CONTROL: Y. Xu,T. Kanade
ISBN: 0-7923-9265-5
NEURAL NETWORKS
IN ROBOTICS
Edited by:
George A. Bekey
Professor, Dept. of Computer Science
University of Southern California
Kenneth Y. Goldberg
Asst. Professor, Dept. of Computer Science
University of Southern California
Most of the papers contained in this book were presented at the
First Workshop on Neural Networks in Robotics,
sponsored by
University of Southern California's
Center for Neural Engineering
October 1991
SPRINGER SCIENCE+BUSINESS MEDIA, L L C
I
Library of Congress Cataloglng-In-Publicatlon Data
Neural networks in robotics / edited by George A. Bekey, Kenneth Y. Goldberg.
p.em.
"Most of the papers contained in this book were presented at the First
Workshop on Neural Networks in Roboties, sponsored by University of
Southern California's Center forNeural Engineering, October 1991."
Includes bibliographical references and index.
ISBN 978-1-4613-6394-1 ISBN 978-1-4615-3180-7 (eBook)
DOI 10.1007/978-1-4615-3180-7
1. Robots--Control systems--Congresses. 2. Neural networks (Computer
scienee)--Congresses. 1. Bekey, George A., 1928- . II. Goldberg,
Kenneth Y. III. Workshop on Neural Networks in Robotics (1st : 1991 : 1.0;
Angeles, Calif.) IV. University of Southern California. Center for N e ura 1
Engineering.
TJ211.35.N48 1992
629.8'92--de20 92-27162
CIP
Copyright © 1993 by Springer Science+Business Media New York
Originally published by Kluwer Academic Publishers in 1993
Softcover reprint of the hardcover 1st edition 1993
AII rights reserved. No part of this publication may be reproduced, stored in a retrieval
system or transmitted in any form or by any means, mechanical, photo-copying, record ing,
or otherwise, without the prior written permission of the publisher, Springer Science+
Business Media, LLC.
Printed on acid-free paper.
Table of Contents
Foreword. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. vii
Introduction by Andrew Barto . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. ix
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. xii
Section I TRAJECTORY GENERATION ....................... 1
1. Learning Global Topological Properties of Robot Kinematic Mappings for
Neural Network·based Configuration Control:
DeMers, Kreutz-Delgado .................................. 3
2. A One-eyed Self Learning Robot Manipulator: Krose,
van der Smagt, Groen ...............•................... 19
3. A CMAC Neural Network for the Kinematic Control of Walking Machine:
Un, Song ............................................ 29
4. Neurocontroller Selective Learning from Man-in-the.Loop Feedback
Control Actions: Kuan, Whittle, Bavarian ..................... 45
5. Application of Self· Organizing Neural Networks for Mobile Robot
Environment: Najand, Lo, Bavarian ......................... 85
6. A Neural Network Based Inverse Kinematics Solution in Robotics:
Li, Zeng ............................................. 97
7. Hopefield Net Generation and Encoding of Trajectories in Contained •
Environment: Bersini, Soreling, Decossaux ..... . . . . . . . . . . . . . . . 113
Section II - RECURRENT NETWORKS ........................ 129
8. Some Preliminary Comparisons Between a Neural Adaptive Controller
and a Model Reference Adaptive Controller: Saerens, Sofuet, Renders,
Bersini ............................................. 131
9. Stable Nonlinear System Identification Using Neural Network Models:
Polycarpou, Ioannou ..................................... 147
10. Modeling of Robot Dynamics by Neural Networks with Dynamic
Neurons: Chassiakos, Kosmatopoulos, Christodoulou ............. 165
11. Neural Networks Learning Rules for Control: Uniform Dynamic
Backpropagation, and the Heavy Adaptive Learning Rule: Seube ... 177
12. Parameter Learning and Compliance Control Using Neural Networks:
Venkataraman, Gulati, Barken, Toomarian ..................... 193
13. Generalisation and Extension of Motor Programs for a Sequential
Recurrent Network: Kumaresan, Sharkey ..................... 217
14. Temporally Continuous vs. Clocked Networks: PearImutter ........ 237
Section III - HYBRID CONTROLLERS ........................ 253
15. Fast Sensorimotor Skin Acquisition Based on Rule-Based Training
of Neural Nets: Handelman, Lane ........................... 255
16. Control of Grasping in Robot Hands by Neural Networks and
Expert Systems: Bekey .................................. 271
vi
17. Robot Task Planning Using a Connectionist/Symbolic System:
Moed, Kelley . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
Section IV - SENSING . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
18. Senses, Skills, Reactions and Reflexes Learning Automatic Behaviors
in Multi-sensory Robotic Systems:
Gelfand, Flax, Endres, Lane, Handelman . . . . . . . . . . . . . . . . . . . 319
19. A New Neural Net Approach to Robot 3D Perception and Visuo-
Motion Coordination: Lee . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
20. Connectivity Graphs for Space-Variant Active Vision:
Wallace, Ong, Bederson, Schwartz . . . . . . . . . . . . . . . . . . . . . . . 349
21. Competitive Learning for Color Space Division:
Uchitama, Tomiyasu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
22. Learning to Understand and Control in a World of Events: Elsley . 389
23. Self-selection of Input Stimuli for Improving Performance:
Nolfi, Parisi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
Section V - BIOLOGICAL SYSTEMS . . . . . . . . . . . . . . . . . . . . . . 419
24. A biologically-Inspired Architecture for Reactive
Motor Control: Massone . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
25. Equilibria Dynamics of a Neural Network Model for Opponent
Muscle Control: Bullock, Contreras-Vidal, Grossberg . . . . . . . . . . . 439
26. Developmental Robotics'- A New Approach to the Specification
of Robot Programs: Fagg . . . . . . . . . . . . . . . . . . . . . . . . . . . . 459
27. A Kinematics and Dynamics Robot Control System Based on
Cerbro-cerebellar Interaction Modelling: Allemand, Blanc,
Bumod, Dufosse, Lavayssiere . . . . . . . . . . . . . . . . . . . . . . . . . . 487
28. What Frogs' Brains Tell Robots' Schemas: Arbib . . . . . . . . . . . . 503
29. Modulation of Robotic Motor Synergies Using Reinforcement
Learning Optimization: Lane, Handelman, Gelfand . . . . . . . . . . . . 521
30. Using Optimal Control to Model Trajectory Formation and
Perturbation Response in a Prehension Task: Hoff . . . . . . . . . . . 539
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 559
Foreword
This book addresses the intersection of the fields of Robotics and Neural Net-
works, both natural and artificial. Most of the papers contained in it were
presented at the First Workshop on Neural Networks in Robotics, sponsored
by USC's Center for Neural Engineering in October 1991.
Robots are physical devices constructed to imitate aspects of human or animal
behavior which involve interaction with the world, such as manipulation or
locomotion. Roboticists try to build systems which can integrate perceptual
inputs smoothly with motor responses, even in the presence of novel stimuli
and changes in the environment. The ability of living systems to learn and to
adapt provides the standard against which robotic systems are judged.
In order to emulate these abilities, a number of investigators have attempted
to create robot controllers which are modeled on known processes in the brain
and musculo-skeletal system. Such computational models have two purposes:
they form the basis for biologically inspired approaches to the design of robot
control, while at the same time they yield some insight into the behavior of the
brain mechanisms they represent.
From an applied point of view, artificial neural networks may be viewed simply
as a wide class of computational models for representing input/output relations,
where input signals propagate through weighted connections to a network of
sum and threshold processors. In this book the term "connectionist" will be
used as a synonym for methods that employ artificial neural networks. The
weights can be obtained by training from examples rather than by explicit pro-
gramming. From this point of view, such networks are also extremely attractive
for robot control. The solution of the inverse kinematic and dynamic equations
for robots is complex and computationally intensive. Furthermore, robot pa-
rameters (such as physical dimensions, moments of inertia or joint friction) are
not known precisely. Hence, it is appealing to consider a method of compu-
tation which, based on measured examples, could yield the relation between
robot end effector coordinates and the corresponding joint variables.
There are historical parallels between the study of neural networks and robots.
Both are biologically inspired in their origin, and both have acquired important
engineering applications in their own right. This book is dedicated to exploring
their intersections, their applications and their limitations. The papers in this
viii
volume are dedicated to answering such questions as: What can neural networks
contribute to improvement of robot control? How faithful a model of biological
processing is needed for good integration of perception and action in robotic
systems? Are there advantages in integrating neural networks with knowledge-
based systems in the design of robot controllers?
An Introduction by Andrew Barto provides a context for this collection, which
is divided into five sections. Section 1 addresses problems related to generat-
ing collision-free trajectories. Section 2 focuses on recurrent networks, where
feedback channels are introduced to allow neural networks to more accurately
model dynamical systems. The papers in Section 3 explore hybrid models that
combine neural networks with expert systems for planning and control. Section
4 focuses on neural methods for treating sensor data, such as camera inputs. Fi-
nally, in Section 5 we consider approaches to robot control which are explicitly
based on models of brain processes.
Undoubtedly, an aura of "hype" has compromised the study of both neural
networks and robotics in recent years. Exaggerated claims of actual and po-
tential accomplishments have affected the credibility of both fields. When the
two subjects are combined, as in this book, the danger is compounded. Yet,
neural net architectures hold promise in many aspects of motor control and
sensor integration. We feel it is the responsibility of all workers in the field to
clarify realistic expectations to ground their results in the context of existing
theory. Indeed, as Polycarpou and Iannou state in this volume, "The construc-
tion of accurate mathematical models from observations of physical activity is
an integral part of the general scientific methodology" .
It is natural to look at the brain for prototypes of information processing, where
input data are seamlessly integrated to control a variety of actuators while
adapting to changing environments. As Arbib says in this book, "What makes
this of interest to roboticists is that it exemplifies a new style of computing
that will become of increasing importance as robot control is distributed over
variegated processors for sensing and motor control". We hope that the papers
collected here will contribute both to scientific methodology and to new styles
of computation.
George A. Bekey
Kenneth Y. Goldberg
Los Angeles, California
Artificial Intelligence, Neural Networks, and
Control
Andrew G. Barto
University of Massachusettes
Before the term Artificial Intelligence (AI) was coined in the early 1960s, the
control of dynamical systems was providing a significant common ground for
researchers studying natural and artificial systems. Cyberneticians were explic-
itly concerned with how control principles could help elucidate the behavior of
organisms. Although control theory contributed to the setting from which AI
emerged (Arbib, 1987), its importance was overshadowed by AI's focus on men-
tal faculties, such as reasoning, that are more readily associated with intelligence
than is the control of respiration or movement. AI progressed by an almost ex-
clusive attention to computational tasks that could be performed off-line based
on perfect information about the problem being solved. The essence of con-
trol, on the other hand, is on-line performance in the presence of uncertainty.
More recently, however, interest is increasing in AI systems that are situated,
or embedded, in real environments in which uncertainty and time contstraints
playa role (e.g., Dean and Wellman, 1991). Control principles are emerging
as important adjuncts to AI methods. At the same time, researchers studying
artificial neural networks are focusing on control as a promising domain for
practical applications as well as for its potential to further our understanding
of biological control systems.
According to a somewhat idealized view of the history of artificial neural net-
work research in relation to engineering and AI (Barto, 1990), research diverged
in the early 1960s to form branches associated with engineering, on the one
hand, and sequential symbol manipulation, or AI, on the other. This marked
the end of what Minsky and Papert (1988) called the "romantic period" during
which biologically inspired computational schemes were first studied. Signifi-
cantly, the wave of research in artificial neural networks that we are now experi-
encing emerged more from the AI tradition than from the engineering tradition.
Although most neural network algorithms are very closely related to engineer-
ing methods, network research tends to continue the AI tradition of addressing
problems that are not formulated with a high degree of mathematical structure.
Rather than focusing on rigorous mathematical theory, AI researchers focus on
heuristic principles that can be applied when the assumptions of rigorous theo-
ries are far from being satisfied. Artificial neural network research tends toward
x
an experimental approaeh more like that of AI than of the engineering disci-
plines to which, technically, network methods have the closest ties. Artificial
neural network research is therefore providing a much needed bridge between
lines of research that diverged thirty years ago.
The idea of memory intensive computation provides a unifying perspective on
the artificial neural network research most relevant to control. According to this
perspective, networks designed for associative memory, pattern classification,
and function approximation are memory devices that can replace sequential
computation with memory access. Instead of computing results from input
data by using a time-consuming sequential computation, one can pre-compute
these results for a range of possible input data and store them in some fashion
that permits rapid access. Input data then serve as indices for memory recall
instead of grist for a sequential computation. Storing information in a network's
connection weights via a learning process is a more general version of storing
information in a lookup table's memory locations. Accessing information from a
network is a relatively shallow computation that can be accomplished in roughly
the same amount of time for all input data; it can replace a deep sequential
computation whose duration varies with the input data. Consequently, artificial
neural networks, as well as other kinds of memory intensive systems, should
have great utility in applications such as control, where real-time operation is
essential. Furthermore, the remarkable hardware advances of the last thirty
years have made memory intensive approaches to control quite feasible.
Bias is the term used to describe a learning system's predisposition for learning
some things at the expense of others. By having varying degrees of built-in
structure, networks can fall almost anywhere in the continuum from unbiased
to highly-biased memory systems. Lookup tables are near the unbiased end
of this continuum because they do not impose constraints other than a cer-
tain grain 'of quantization on the data they store. Memory systems nearer the
highly-biased end assume specific functional relationships between their inputs
and outputs. Highly-biased memory systems generally have far fewer degrees-
of-freedom than lookup tables, and the form of their bias provides the basis
for generalizing beyond the data with which they have direct experience (which
will be successful to the extent that the bias matches the functional relation-
ship actually present in the data). According to this perspective, conventional
adaptive control methods fall near the highly-biased extreme of this contin-
uum. Networks, along with other flexible memory systems, can strike a balance
between the unavoidable tradeoffs in a way not possible for systems that are
either extremely biased or extremely unbiased.
One avenue for applying networks to control is therefore to replace the ap-
propriate components of a conventional adaptive control system by adaptive
networks. Most applications of artificial neural networks to adaptive control
reported in the literature are the result of such substitutions. Although the
theorems proven for conventional methods do not readily extend to most neural
network controllers, this approach clearly shows how memory intensive meth-
ods can extend conventional adaptive control methods. Despite the successes
reported for this approach, the most dramatic applications of artificial neural
xi
networks to control will not necessarily be the result of such conservative ex-
tensions of conventional adaptive control methods. Less conservative methods
may have advantages in problems that so blatantly violate conditions required
for theoretical guarantees that a control engineer would never consider them
as candidates for adaptive control. Because some of the control problems that
animals routinely learn to solve fall into this category, conventional adaptive
control methods can be misleading guides to thinking about control in biological
systems. It seems unlikely that ancestral organisms ever faced linear environ-
ments with Gaussian uncertainty, or that an ability to control linear systems
was an evolutionary prerequisite for controlling nonlinear systems. Perhaps a
better understanding of how animals learn to control highly nonlinear systems
under great uncertainty can provide guidance for the development of similarly
effective engineering methods.
M. A. Arbib, Brains, Machines, and Mathematics, Second Edition. New York:
Springer-Verlag, 1987.
A. G. Barto, Connectionist Learning for Control: An Overview, In T. Miller
and R. S. Sutton and P. J. Werbos (eds.), Neural Networks for Control, pp. 5-58.
Cambridge, MA: The MIT Press, 1990.
T. L. Dean and M. P. Wellman, Planning and Control. San Mateo, CA: Morgan
Kaufmann, 1991.
M. L. Minsky and S. A. Papert, Perceptrons: An Introduction to Computational
Geometry, Expanded Edition. Cambridge, MA: The MIT Press, 1988.
Acknowledgements
The editors would like to thank Michael Arbib, Director of USC's Center for
Neural Engineering, for hosting this workshop and for serving on the program
committee. We would also like to thank Kusum Shori of USC for her excel-
lent secretarial assistance, and Tom Stone and Stephanie Faulkner of Kluwer
Academic Publishers for their help in realizing this manuscript.
