INTERLEAVING PLANNING AND EXECUTION FOR AUTONOMOUS ROBOTS

Transcription

1 INTERLEAVING PLANNING AND EXECUTION FOR AUTONOMOUS ROBOTS

2 THE KLUWER INTERNATIONAL SERIES IN ENGINEERING AND COMPUTER SCIENCE ROBOTICS: VISION, MANIPULATION AND SENSORS Consulting Editor Takeo Kanade Other books in the series: GENETIC LEARNING FOR ADAPTIVE IMAGE SEGMENTATION, Bir Bhanu, Sungkee Lee ISBN: SPACE-SCALE THEORY IN EARLY VISION, Tony Lindeberg ISBN NEURAL NE1WORK PERCEYl10N FOR MOBILE ROBOT GUIDANCE, Dean A. Pomerleau ISBN: DIRECTED SONAR SENSING FOR MOBILE ROBOT NA VIGATION,Iohn 1. Leonard, Hugh F. Durrant-Whyte ISBN: A GENERAL MODEL OF LEGGED WCOMOTION ON NATURAL TERRAINE, David 1. Manko ISBN: INTELLIGENT ROBOTIC SYSTEMS: THEORY, DESIGN AND APPLICATIONS, K. Valavanis, G. Saridis ISBN: QUALITATIVE MOTION UNDERSTANDING, W. Burger, B. Bhanu ISBN: NONHOWNOMIC MOTION PLANNING, Zexiang Li, I.F. Canny ISBN: SPACE ROBOTICS: DYNAMICS AND CONTROL, Yang sheng Xu, Takeo Kanade ISBN: NEURAL NE1WORKS IN ROBOTICS, George Bekey, Ken Goldberg ISBN: X EFFICIENT DYNAMIC SIMULATION OF ROBOTIC MECHANISMS, Kathryn W. Lilly ISBN: MEASUREMENT OF IMAGE VELOCITY, David 1. Fleet ISBN: INTELLIGENT ROBOTIC SYSTEMS FOR SPACE EXPLORATION, Alan A. Desrochers ISBN: COMPUTER AIDED MECHANICAL ASSEMBLY PLANNING, L. Homen de Mello, S. Lee ISBN: PERTURBATION TECHNIQUES FOR FLEXIBLE MANIPULATORS, A. Fraser, R. W. Daniel ISBN: DYNAMIC ANALYSIS OF ROBOT MANUPULATORS: A Cartesian Tensor Approach, C. A. Balafoutis, R. V. Patel ISBN:

3 INTERLEA VING PLANNING AND EXECUTION FOR AUTONOMOUS ROBOTS by fiah Reza Nourbakhsh Stanford University.., ~ SPRINGER SCIENCE+BUSINESS MEDIA, LLC

4 ISBN ISBN (ebook) DOI / Library of Congress Cataloging-in-Publication Data A C.I.P. Catalogue record for this book is available from the Library of Congress. Copyright ~ 1997 by Springer Science+Business Media New York Origina1ly published by Kluwer Academic Publishers in 1997 Softcover reprint of the hardcover 1 st edition 1997 AlI 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, recording, or otherwise, without the prior written permission of the publisher, Springer Science+Business Media, LLC. Printed on acid-free paper.

5 to Noah and Zachary

6 Contents Preface Acknowledgments xiii xv 1 Introduction Motivation Contributions Contents of Chapters 8 2 Perception and Action The Situated Robot and its Environment Defining State and State Update Comparison to other Conventional Approaches Conclusion 19 3 Formalizing Incomplete Information Sources of Incomplete Information Representing Incomplete Information State-Set Tracking: Execution Monitoring Related Work 33 4 Goal-Directed Control Systems Defining a Problem Instance Advance Planning Control Systems Related Work 51 5 Interleaving Planning and Execution Introduction: Interleaving Premature Search Tennination Experimental Results Related Work 63

7 viii IIlah R. Nourbakhsh 6 Using Assumptions to Oversimplify Introduction Assumptions Assumptive Control Systems Continuous Selection Algorithm The Cost of Assumptions Large-scale Real-World Experimental Results Related Work Conclusion 94 7 Strategic Subgoaling: Using Abstraction Systems Introduction Problem Spaces Abstraction 1, Examples of Abstraction Systems and their Cost Related Work and Discussion Generalizing beyond State Sets Introduction Representation Examples Conclusion Conclusions 131 Bibliography 137 Index 143

8 List of Tables 4.1 Search space size for five variations of two simulated problems Running times for cond-plan without useless plan pruning (CPA) and 51 with pruning (CPAb) 5.1 The size of the search space for seven problems The running times and execution lengths for each problem A comparison of the planning cost of advance planning versus 83 Assumptive planning for three domain classes 6.2 A comparison of the conditional and Assumptive search space sizes 88 for the CS224 robot problem of figure A comparison of the conditional and Assumptive search spaces for 92 the final contest at the AAAl1994 National Robot Competition

9 List of Illustrations 1.1 A comparison of advance planning and interleaving search space A birds-eye view of a robot lost in a maze A comparison of advance planning and Assumptive planning A comparison of advance planning and an abstraction system 7 2.l A situated robot's actions and percepts A finite automata view of Figure The Mazeworld domain A partial world state graph for the Mazeworld domain of Fig l An example of model error A robot with an unreliable move action A robot with an inconsistent percept The state set representation of knowledge about world state Graphical examples of effect Effectory collapse Perceptual collapse Effectory expansion Perceptual unreliability Perceptual aliasing A partial state-set graph for a Mazeworld Goal sets A conditional plan for changing highway lanes A behavioral graph for a Mazeworld State-set overlay A flowchart depicting the CPA control system An overhead view of the MJH domain An example Wumpus World game A flowchart depicting the DPA control system Search horizon example An example of a forced move 59

10 xii R. Nourbakhsh 5.4 Conditional planning with uncertain initial conditions The search space of an Assumptive system An assumption that reduces initial state set size A Mazeworld example with initial uncertainty The dangers of a deadly snake pit A demonstration of nondeterministic move An example of perceptual branching with false negatives A flowchart depicting the basic Assumptive Algorithm An example of state-set unreachability An example of false goal detection An example of assumptive overcommitment A flowchart depicting Continuous Selection Algorithm 6.12 An example Mazeworld in which Indiana Robot faces snake pits The CS 224 Final Contest maze Three CS 224 students at work The CS 224 Assumptive examination Geometric and topological maps of Dervish Dervish in its final configuration Quantization of a geometric map A comparison of a bipartite conditional graph and a hyperarc graph A state graph-based example of abstraction Abstract and ground conditional plan fringes An example of useful abstraction An example of abstraction systems A flowchart depicting the problem space control system An example run of a three-tier abstraction system The horizon problem of geometric abstraction Hyperarcs in an abstraction space Hyperarcs demonstrating continent abstraction Path nonoptimality example TC-actions An example of the property set representation An example of polygonal position representation for mobile robots A robot navigation problem solved using the Viable Plan Heuristic 129

11 Preface An unfortunate gap exists in the robotics community between theory and practice. Nowhere has this diconnect been more apparent than in the area of robot architecture design. Consequently, many research results have been difficult, if not impossible, to reproduce because assumptions, architectural decisions and conditions requisite to success are poorly specified. The answer is a good theoretical foundation that is directly relevant to the robot design problem. Good theory can offer design tools that reduce the learning curve, making the real-world design of a robot easier. This book is a modest step in this direction, offering a formal representation for robot architecture, which combines sensing, acting and planning to form an integrated whole. This work also discusses interleaving planning and execution in depth. Interleaving is a necessity in real-world robots that operate with incomplete information. Interleaving systems that preserve the robot's functionality are at the core of this work, and the representation we propose allows us to prove soundness, completeness and optimality results for robot architectures under well specified conditions. After presenting the formal framework in the first four chapters, we describe experimental results on real-world robots. Experimental validation not only needs to show that the theory has been used to drive the implementation of a successful robot, it must also demonstrate that the theory has allowed others to solve robot problems more easily. To this end, Chapter 6 offers a detailed account of CS 224: Introduction to Robot Programming Laboratory. This popular lab class at Stanford University, taught by Professor Michael Genesereth and the author, has given undergraduate and graduate students a chance to apply theoretical principles of robot architecture and interleaving to capable teams of real-world robots. In addition, Chapter 6 describes the robot, Dervish, which used the Assumptive Architecture presented here to win the 1994 AAAI National Robot Contest. Its

12 xiv liiah R. Nourbakhsh long tenn reliability, as well as the simplicity of its control system, bear witness to the usefulness of a theoretical paradigm that can guide architecture design. Chapter 7 demonstrates an important extension of this theoretical frameworlc to cover abstraction, a necessary tool for problem refonnulation in the robotics community. We hope that this book engages real-world robot designers, offering them theoretical tools that make the process of architecture design more well-defined. At the same time, we hope that theoreticians will find this book to be a starting point for future research into the area of robot architecture. This very new discipline contains many more questions than answers.

13 Acknowledgmen ts This text is my most focused educational effort to date. Therefore it is appropriate to begin by acknowledging those who have chosen to devote their lives to education. Kevin Madden was my high school physics teacher and a friend. My daily conversations with him after lunch allowed me to sample his immense knowledge about the natural world. He made the physical world intuitive for his students. Donald Adams taught me creativity. For four years, I spent two hours after school every day in the pottery studio, learning much more than pottery from one with an endless supply of kindness, friendship and humor. Fred White was my high school chemistry teacher. He taught me chemistry during the school years and sailing during the summers. I had almost daily contact with him for seven years and I cannot begin to describe how much this Renaissance Man and good friend has taught me. Fast-forward four years and I acknowledge the one teacher who has the greatest impact on me in the field of Artificial Intelligence. Michael Genesereth convinced me unwittingly to major in Computer Science when I was an undergraduate at Stanford. His class, Discrete Systems, showed me how little I knew about a most exciting field of endeavor. As my undergraduate advisor, graduate advisor, flying companion and good friend, he has taught me much more than Artificial Intelligence. There are two lessons that I value the most; two traits of Professor Genesereth's personality that teach by example. First, he exemplifies lucid science. He formulates his intellectual statements cleanly and precisely, with a solid foundation of formal logic at the core. I will always aspire to his style of science. But the most important lesson I learn from Professor Genesereth concerns his underlying motivation behind doing science: to make the world a better place. It is this consonance between the ethical outlook of my advisor and my own world view that has made the last four years of graduate studies worthwhile. I am certain that I will Ie am from Professor Genesereth for many years to come.

14 xvi liiah R. Nourbakhsh There are those friends who have relentlessly volunteered their time to read my paper drafts and hear my presentations. Their comments have shaped the way I think and the way I write: David Andre, Lise Getoor, Steve Ketchpel, Ofer Matan and many others deserve acknowledgment. I would also like to specifically thank those who have read and commented on this work throughout its various life forms: David Andre, Michael Genesereth, Steve Ketchpel, Jean-Claude Latombe, Matt Mason, Stan Rosenschein. My family has given me the love that sets my research in proper perspective, from my grandparents down through my parents and uncles, all the way, via my brother and his wife, to Noah and Zachary, my two year old nephews whose pictures adorn the office door. But there are newer members of my family that I must also acknowledge. The Barhoums have taken me into their family since my move from Kansas City to Stanford. Julie is an even newer member of family to me; she is one with whom I share another, very special kind of love.

15 INTERLEAVING PLANNING AND EXECUTION FOR AUTONOMOUS ROBOTS