Intelligent Control of Robotic Systems

Transcription

1 Intelligent Control of Robotic Systems

2 International Series on MICROPROCESSOR-BASED AND INTELLIGENT SYSTEMS ENGINEERING VOLUME 25 Editor Professor S. G. Tzafestas, National Technical University of Athens, Greece Editorial Advisory Board Professor C. S. Chen, University of Akron, Ohio, U.S.A. Professor T. Fokuda, Nagoya University, Japan Professor F. Harashima, University of Tokyo, Tokyo, Japan Professor G. Schmidt, Technical University of Munich, Germany Professor N. K. Sinha, McMaster University, Hamilton, Ontario, Canada Professor D. Tabak, George Mason University, Fairfax, Virginia, U.S.A. Professor K. Valavanis, University of Southern Louisiana, Lafayette, U.S.A. The titles published in this series are listed at the end of this volume.

3 Intelligent Control of Robotic SysteIlls by DUSKO KATIC Robotics Laboratory, Mihailo Pup in Institute Belgrade, Serbia, Serbia & Montenegro and MIOMIR VUKOBRATOVIC Robotics Laboratory, Mihailo Pupin Institute Belgrade,Serbia, Serbia & Montenegro Springer-Science+Business Media, B.V.

4 A c.i.p. Catalogue record for this book is available from the Library of Congress. ISBN ISBN (ebook) DOl / Printed on acid-free paper All Rights Reserved 2003 Springer Science+Business Media Dordrecht Originally published by Kluwer Academic Publishers in Softcover reprint of the hardcover I st edition 2003 No part of this work may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, microfilming, recording or otherwise, without written permission from the Publisher, with the exception of any material supplied specifically for the purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the work.

5 Contents List of Figures List of Tables Preface Acknow ledgments 1. INTELLIGENT CONTROL IN CONTEMPORARY ROBOTICS 1 Introduction 2 Role and Application of Learning in Robotics 2.1 Introduction to robot learning 2.2 Application of robot learning ix xiii xv XIX NEURAL NETWORK APPROACH IN ROBOTICS Introduction Connectionist Models with Application in Robotics Learning Principles and Learning Rules Applied in Robotics Neural Network Issues in Robotics 4.1 Kinematic robot learning by neural networks 4.2 Dynamic robot learning by neural networks Sensor-based robot learning by neural networks 50 5 Efficient Learning Control for Manipulation Robots by Feedforward Neural Networks Efficient connectionist learning structures Case Study FUZZY LOGIC APPROACH IN ROBOTICS 71 2 Introduction Mathematical foundations 2.1 Fuzzy sets v

6 VI INTELLIGENT CONTROL OF ROBOTIC SYSTEMS 2.2 Operations on fuzzy sets Fuzzy relations Fuzzy logic 79 3 Fuzzy controller Condition interface Fuzzy set definition base Control rules Inference mechanism Action interface 90 4 Direct applications 92 5 Synthesis of fuzzy controller - example from robotics 93 6 Fuzzy algorithms in robotics GENETIC ALGORITHMS IN ROBOTICS Introduction ll3 1.1 Synthesis of GA - example from robotics GAs in Robotics HYBRID INTELLIGENT APPROACHES IN ROBOTICS l33 1 Basic Ideas of Neuro-Fuzzy Approach Neuro-Fuzzy Algorithms in Robotics Hybrid genetic approaches in robotics SYNTHESIS OF CONNECTIONIST CONTROL ALGORITHMS FOR ROBOT CONTACT TASKS Introduction Fundamentals of Connectionist Control Synthesis Model of the robot interacting with dynamic environment - task setting Synthesis of non-learning control algorithms for robotic contact tasks Synthesis of non-learning algorithms stabilizing robot motion Synthesis of non-learning algorithms stabilizing the interaction force Synthesis of non-learning impedance control as a specific case of unified position/force control Synthesis of Learning Stabilizing Control Laws by Connectionist Structures for Contact Robotic Tasks Connectionist structure and on-line learning algorithm 175

7 Contents Vll 4 Synthesis of Connectionist Learning Impedance Laws for Robotic Contact Tasks Recurrent connectionist structure and on-line learning rules Case Study SYNTHESIS OF COMPREHENSIVE CONNECTIONIST CONTROL ALGORITHMS FOR CONTACT TASKS 193 Introduction Factors Affecting Task Performance and Stability in Robotic Compliance Control Comprehensive Connectionist Control Algorithm Based on Learning and Classification for Compliance Robotic Tasks Acquisition process for classification - the first phase Pure neural network classifier Wavelet network classifier 3.2 On-line compliance control algorithms for contact tasks with environment classification - the second phase Case Study EXAMPLES OF INTELLIGENT TECHNIQUES FOR ROBOTIC APPLICATIONS Introduction Connectionist Reactive Control for Robotic Assembly Tasks by Soft Sensored Grippers Analysis of the assembly process with soft fingers Assembly process Learning compliance methodology by neural networks Experimental results Genetic-Connectionist Algorithm For Compliant Robotic Tasks GA tuning of PI force feedback gains Simulation experiments involving genetic algorithms Connectionist Compensator For Advanced Integrated Vehicle Controller 4.1 Model of the vehicle dynamics 4.2 Control strategy 4.3 Synthesis of the vehicle dynamic controller 4.4 Synthesis of the supplementary neuro-compensator 4.5 Simulation experiments

8 viii References Index About the Authors INTELLIGENT CONTROL OF ROBOTIC SYSTEMS

9 List of Figures 1.1 Hierarchical intelligent robotic systems Multilayer perceptron Radial basis function network Structure of CMAC network Hopfield network ART network SOFM network Specialized learning architecture Generalized learning architecture Feedback-error learning architecture Reinforcement learning architecture Sensor-motor circular reaction Sensor-based learning architecture The UNH biped walking Block diagram of overall biped control system High - level control architecture Comparison of BP and RLS algorithm Comparison of BP and EKF algorithm Position error - comparison of BP and RLS algorithms Position error - comparison of BP and EKF algorithms Position error - with learning and without learning Estimated number of commercial applications of fuzzy systems Membership functions of conventional and fuzzy sets Examples of fuzzy sets 75 IX

10 x INTELLIGENT CONTROL OF ROBOTIC SYSTEMS 3.4 Standard operations on fuzzy sets Examples of linguistic modifiers Components of fuzzy controller Primary fuzzy sets Operational rules Membership functions for 5 linguistic sets 93 3.lO Inference engine and defuzzification of robot controller Fuzzy robot control structure as part of closed loop systems with feedback Tuning fuzzy robot control architecture 97 3.l3 Self-organizing controller Hybrid control scheme lo Fuzzy scheme of force control lo Architecture of the reinforcement learning based gait synthesizer Block diagram of the GA optimization process Structure of neuro-fuzzy network l Neuro-fuzzy architecture for mobile robot navigation in an uncertain working environment 5.3 Determination of reflex force gain using fuzzy and neuro approach for bilateral teleoperation 5.4 Intelligent planner for neuro-fuzzy control of robot machining operations Neuro - GA approach for optimization of robot swing motion Block diagram of stabilization biped control Fuzzy-GA training hierarchical control structure Transfer of human skills to robot controllers by neural network approach Neural network used for peg-in-hole insertion Scheme of learning connectionist law stabilizing robot motion Scheme of learning connectionist law stabilizing interaction force Neural network architecture for compensation of system uncertainties Scheme of learning impedance connectionist law Industrial robot MANUTEC r External error!:1z - position stabilization 186 l37 l39

11 List of Figures Xl 6.9 Error of nonnal force - position stabilization External error b..z - force stabilization Error of nonnal force - force stabilization Non-learning and learning impedance control class I - external error b..x Non-learning and learning impedance control class I - external error b..y Non-learning and learning impedance control class I - external error b..z Non-learning and learning impedance control class I - error of nonnal force Non-learning and learning impedance control class II - external error b..x Non-learning and learning impedance control class II -external error b.. Y Non-learning and learning impedance control class II - external error b..z Non-learning and learning impedance control class II - error of nonnal force Wavelet network classifier Scheme of learning control law stabilizing robot motion with neural classifier Scheme of control law stabilizing interaction force with neural classifier Scheme of learning control law with wavelet network classifier Force error - control law stabilizing robot motion - first set of feedback gains Internal error - control law stabilizing robot motion - first set of feedback gains Force error - control law stabilizing robot motion - second set of feedback gains Internal error - control law stabilizing robot motion - second set of feedback gains Steady-state process for nonnal force - control law stabilizing robot motion - second set of feedback gains Square criterion during learning epochs Comparison of desired and real output of network Nonnal force with neural classifier - stabilizing robot motion 218

12 xii INTELLIGENT CONTROL OF ROBOTIC SYSTEMS 7.13 Normal force without neural classifier - stabilizing robot motion Internal error for stabilizing interaction force control algorithm Force error for stabilizing interaction force control algorithm Circular tracking - comparison with and without neural classifier Force error - comparison with and without neural classifier Interaction force - non-learning impedance control algorithm Comparison with and without classifier Be1grade-USC-IIS multifingered hand Model II Experimental fixture Selected compensation directions Structure of multilayer perception used in off-line learning procedure Square error criterion during process of off-line learning Best force feedback gains K F P and K F I according to ISE and ITAE criterions Best values of ISE and ITAE criterion during evaluation process Block-scheme of the road vehicle autopilot with centralized dynamic controller and supplementary neuro-compensator Structure of the learning process of the chosen ANN Error in learning process Accuracy indices in position tracking Accuracy indices in velocity tracking 255

13 List of Tables 2.1 Parameters of Manutec r3 manipulation mechanism Parameters of D.C. actuators for robot MANUTEC r Rule base for fuzzy controller Fuzzy relation for passage width Fuzzy relation for robot velocity Fuzzy relations of the control algorithm Production rules GA evaluation process for the topology of the neural classifier Environment model profiles Inputs and target outputs of neural classifier Stiffness parameters of robot environment models Damping parameters of robot environment models Inertia parameters of robot environment models PD local gains for satisfaction of performance criterion 213 Xlll

14 Preface As robotic systems make their way into standard practice, they have opened the door to a wide spectrum of complex applications. Such applications usually demand that the robots be highly intelligent. Future robots are likely to have greater sensory capabilities, more intelligence, higher levels of manual dexterity, and adequate mobility, compared to humans. In order to ensure high-quality control and performance in robotics, new intelligent control techniques must be developed, which are capable of coping with task complexity, multi-objective decision making, large volumes of perception data and substantial amounts of heuristic information. Hence, the pursuit of intelligent autonomous robotic systems has been a topic of much fascinating research in recent years. On the other hand, as emerging technologies, Soft Computing paradigms consisting of complementary elements of Fuzzy Logic, Neural Computing and Evolutionary Computation are viewed as the most promising methods towards intelligent robotic systems. Due to their strong learning and cognitive ability and good tolerance of uncertainty and imprecision, Soft Computing techniques have found wide application in the area of intelligent control of robotic systems. Our objective in writing the present book was to produce a fair combination of comprehensive report, survey, theoretical background and special research works with appropriate application examples in the area of intelligent control of robotic systems. This book is chiefly focused on theoretical and application aspects of neural networks, fuzzy logic, genetic algorithms and hybrid intelligent techniques in robotics. The specific emphasis in research work is given to the development of efficient learning rules for robotic connectionist training and synthesis of neural learning algorithms for low-level control in the domain of robotic compliance tasks. The book contains several different examples of applications based on neural and hybrid intelligent techniques. The book is organized in the following way. The first chapter gives an introduction to intelligent control together with a presentation of the basic ideas of intelligent control in robotics. Special attention is paid to the role of learning xv

15 xvi INTELLIGENT CONTROL OF ROBOTIC SYSTEMS as one of the main intelligent capabilities of the control algorithms dealt with in this book. The focus then shifts (Chapter 2) to neural networks (connectionist systems), to review the fundamental concepts and learning principles, discus important implementation issues, and provide a survey of connectionist algorithms in contemporary robotics. The connectionist models for robotic purposes are the subject of a special analysis. The issues of neural networks in robotics with the applications in the domain of kinematic, dynamic, and sensor-based learning are analyzed. A special part of this chapter represents the development of new, efficient learning structures for robot training, followed by appropriate simulation examples. The various implementation issues are addressed by design exploration and the verification of intelligent control paradigms for a variety of robotic applications, including industrial, service, mobile, locomotion, space, underwater and other types of robotics. In the following chapters, the fundamentals of fuzzy logic (Chapter 3), genetic algorithms (Chapter 4) and hybrid intelligent technics (Chapter 5) are presented together with a comprehensive report on their importance and recent applications in autonomous robotic systems. The synthesis of new, advanced learning algorithms for robotic contact tasks by nonrecurrent and recurrent connectionist structures is presented in Chapters 6 and 7 as the main research contribution. In Chapter 6, which includes a survey of connectionist algorithms for robotic contact tasks, the main concern is the development of learning control algorithms as an upgrade of conventional non-learning control laws for robotic compliance tasks (algorithms for stabilization of robot motion, stabilization of robot interaction force and impedance algorithms). In view of the important influence of robot environment, a new, comprehensive learning approach, based on simultaneous classification of robot environment and learning of robot uncertainties, is reported in Chapter 7. The proposed comprehensive algorithms include the synthesis and application of two newly proposed classifiers: pure perceptron classifier and wavelet network classifier. Both chapters are accompanied by simulation studies to validate the proposed algorithms. The book concludes with Chapter 8, which presentates some interesting examples of connectionist approaches, together with some supporting intelligent techniques for special robotic applications. The examples include connectionist reactive control of the soft-sensored grippers robotic assembly tasks, a special genetic-connectionist algorithm for compliant robotic tasks and a connectionist approach to robotized road vehicles (automobiles). This book has evolved from many years of research work and teaching in the areas of automatic control, intelligent robotics and robotics in general. There is one unifying aspect to the work reported in the book, namely its interdisciplinary nature, especially in the combination of robotics, control theory, com-

16 PREFACE XVll putational intelligence and soft computing paradigms. It is our hope that it will be useful to a wide audience of engineers, ranging from students and academic researchers, to the practitioners. The presented text can satisfactorily serve as an educational tool for engineering students interested in pursuing the study of intelligent autonomous robotic systems, as well as a starting base for researchers in the ongoing research in these areas. Belgrade, April 2003 DUSKO KATIC, MIOMIR VUKOBRATOVIC

17 Acknowledgments During the years taken to research and write this book, consciously and subconsciously, we have picked up material from a knowledge base called intelligent control in robotics. Hence, we have had the pleasure and privilege of interacting with many researchers throughout the world and to whom we owe our deepest thanks. The long-term support of our research on intelligent control in robotics by the Serbian Ministry of Science and Technology is gratefully acknowledged. We would like to express our great appreciation to Branislav Borovac for our collaborative work on the problems of connectionist reactive control for robotic assembly tasks (Chapter 8). We want to thank Aleksandar Rodic for our collaborative research in the area of neural controllers for robotized vehicles (Chapter 8). We also want to express our gratitude to Branko Karan for his participation in writing some parts of the text about the general theory of fuzzy logic and their application to robotic systems (Chapter 3). Finally, we thank Luka Bjelica who has proofread the final version of the manuscript and thereby removed some of the errors and gave us valuable suggestions for improvements. XIX