DARWIN2K An Evolutionary Approach to Automated Design for Robotics

Transcription

1 DARWIN2K An Evolutionary Approach to Automated Design for Robotics

2 Other books in the series: THE KLUWER INTERNATIONAL SERIES IN ENGINEERING AND COMPUTER SCIENCE ROBOTICS: VISION, MANIPULATION AND SENSORS Consulting Editor Takeo Kanade ENGINEERING APPROACHES TO MECHANICAL AND ROBOTIC DESIGN FOR MINIMALLY INVASIVE SURGERIES A. Faraz, S. Payandeh ISBN: ROBOT FORCE CONTROL B. Siciliano, L. Villani ISBN: DESIGN BY COMPOSmON FOR RAPID PROTOTYPING M.Binnard ISBN: TETROBOT: A Modular Approach to Reconfigurable Parallel Robotic.. G..J. Hamlin, A.C. Sanderson ISBN: INTELLIGENT UNMANNED GROUND VEHICLES: Autonomous Navigation Research at Carnegie Mellon M. Hebert, C. Thorpe, A. Stentz ISBN: INTERLEAVING PLANNING AND EXECUTION FOR AUTONOMOUS ROBOTS IIlah Reza Nourbakhsh ISBN: GENETIC LEARNING FOR ADAPTIVE IMAGE SEGMENTATION Bir Bhann, Sungkee Lee ISBN: SPACE-SCALE THEORY IN EARLY VISION Tony Lindeberg ISBN NEURAL NETWORK PERCEPTION FOR MOBILE ROBOT GUIDANCE Dean A. Pomerleau ISBN: DIRECTED SONAR SENSING FOR MOBILE ROBOT NAVIGATION John J. Leonard, Hugh F. Durrant Whyte ISBN: A GENERAL MODEL OF LEGGED LOCOMOTION ON NATURAL TERRAlNE David J.Manko ISBN: INTELLIGENT ROBOTIC SYSTEMS: THEORY, DESIGN AND APPLICATIONS K. Valavanis, G. Saridis ISBN: QUALITATIVE MOTION UNDERSTANDING W. Burger, B. Bhanu ISBN:

3 DARWIN2K An Evolutionary Approach to Automated Design for Robotics by Chris Leger The Robotics Institute Carnegie Mellon University SPRINGER SCIENCE+BUSINESS MEDIA, LLC

4 Library of Congress Cataloging-in-Publication Data Leger, Chris, Darwin2K: an evolutionary approach to automated design for robotics / by Chris Leger. p. cm.-(kluwer international series in engineering and computer science; SECS 574. Robotics) Includes bibliographical references and index. ISBN ISBN (ebook) DOI / I.Robotics. 2. Computer-aided design. 3. Darwin2K. I. Title. II. Kluwer international series in engineering and computer science; SECS 574. m. Kluwer international series in engineering and computer science. Robotics. TJ211.LA '92--dc Copyright 2000 Springer Science+Business Media New York Origina11y published by Kluwer Academic Publishers, New York in 2000 Softcover reprint of the hardcover 1 st edition 2000 AII rights reserved. No part of this publication may be reproduced, stored in a retrieval system or transrnitted in any forrn or by any means, mechanical, photocopying, recording, or otherwise, without the prior written permission ofthe publisher, Springer Science+Business Media, LLC Printed on acid-free paper.

5 Contents 1 Introduction Related Work Approach Organization of the book 19 2 Representation and Architecture Robot representation Extensible Software Architecture Summary 41 3 Synthesis Methodology Introduction Genetic Operators Selecting Configurations for Reproduction and Deletion Synthesis Process Summary 75 4 Robot Evaluation Simulator Architecture Computing Robot State Singularity-Robust Inverse Controller Robot Dynamics Link Deflection Path Planning Other Capabilities and Components Metrics Summary 132

6 5 Experiments and Demonstration Task 1: A free-flying robot for orbital maintenance Task 2: A fixed-base manipulator Task 3: Simplified manipulation task for characterizing synthesizer performance Task 4: A Material-handling robot Task 5: An antenna-pointing mechanism Task 6: A walking robot for space trusses Discussion Conclusion Lessons Learned Future Directions 230 Appendix A: OOP and Class Hierarchy Appendix B: Module descriptions Appendix C: Detailed robot descriptions Glossary References Index vi

7 The software described in this book is freely available at:

8 Acknowledgments This book is based on my PhD dissertation, and it is thus appropriate to mention the numerous people who helped me complete that work. My family has provided constant encouragement throughout my academic career; I don't recall any of them ever asking questions like "So, when are you finally going to get a job?" John Bares provided an optimal amount of guidance for my research and was a good source of motivation -- sometimes he seemed more excited about my work than I was. My thesis committee (Manuela Veloso, Rob Ambrose, Pradeep Khosla, and Chris Pared is) kept me on track in the early stages of my thesis and challenged me to think of the bigger picture during my defense; some of their comments have been incorporated into Chapters 5 and 6. Whether he intended to or not, Andrew Johnson acted as a big brother of sorts and is partly to blame for my decision to apply to grad school in the first place. I'm a big proponent of finding balance in life, and I have my friends to thank for helping me achieve that. My band-mates -- Kevin Lenzo, Matt Siegler, Garth Zeglin, Scott Baugh, Scratchy, and Martin Martin (all of whom have had the Grad Student Experience) -- gave me an much-needed outlet for creative musical energies. Jorgen Pedersen introduced me to rock climbing, which has since become an obsession in my life. Chances are I would have gone nuts without having climbing and my climbing partners to keep me sane, though I suppose most people think climbers are nuts anyway. Jorgen also wrote the RTC communications package used by Darwin2K. John and Lee Bares kindly let me use their woodshop for other stress-relieving extracurricular activities. Jeff Smith, Parag Batavia, and my other classmates also helped me in the pursuit of slack. Jason Bannister created the cover art. Finally, I'd like to thank Rosie Jones for friendship and support, for proof-reading the original dissertation, and for suggesting both the name "Darwin2000" and the final title of this book.

9 Foreword Since its early beginnings a quarter century ago, robotics has lacked a formal toolset to guide the selection and arrangement of degrees-of-freedom and linkages, a process called configuration. A robot designer is faced with the dilemma of a nearly endless set of design alternatives: Should a robotic device have three revolute links followed by an extensional link or two revolute-extensional link sequences? Will the forces on a manipulator during a high speed motion cause unacceptable tip deflection anywhere within the task trajectory? Might the addition of redundant degree-of-freedom offer a power saving alternative and improve tip forces in certain regions of the work space? Will the maximum manipulator payload cause instability of a wheeled base when operating on an incline? Until now, the robot designer has not a means to rapidly explore these kinds of configuration questions. Typically, configuration occurs quickly or is skipped altogether in favor of the detailed design stage which works to finalize dimensions, materials, gear ratios, connections and the like. Development teams are taxed to simply create a "viable" solution to a set of task requirements, let alone imagine creating an "optimized" solution. However, it is during configuration that many of the decisions occur which characterize the complexity, reliability, efficiency and overall quality of a robotic device. The best detailed design cannot significantly improve upon a poor or vastly suboptimal configuration. With the creation of Darwin2K, Chris Leger offers the robot designer a powerful tool that can greatly improve the quality of a robot configuration. Darwin2K efficiently represents complex robot systems and utilizes an evolutionary algorithm to create thousands of configuration alternatives, each of which is analyzed and simulated with regard to the task requirements. Using genetic operators such as crossover and mutation, new populations are generated from successful configurations and the process continues to evolve successively improved designs. Realizing that task description is critical in constrain-

10 ing a robot configuration, Darwin2K even optimizes selected task parameters to improve overall robot performance. Armed with Darwin2K, a robot designer can focus on guiding the configuration process and not be troubled with the exhaustive cycle of generating and testing a wide variety of possible candidate configurations. In a matter of several days, a designer can input a task description into Darwin2K and guide the program through examination and testing of hundreds of thousands of alternative configurations. The final configuration or small set of configurations that results then serves as the step-off point for manual design and component sizing. My own experience in designing a dozen robots is that the configuration process is ad hoc and is quite often closely related to the designer's judgement and the amount of time available. A common result is that the first viable configuration remains as the "one to beat." Usually it persists as the "best" configuration simply because it was derived early on and project resources simply limit the creation of many, if any, competitive alternatives. Consider for instance a wheeled mobile robot that we designed for demolition in a crippled nuclear power plant. I would say that we created no more than three or four viable configurations before the project schedule demanded configuration closure and final design. If I could apply Darwin2K to that same problem now, several hundred thousand alternatives would be examined, with thousands of viable configurations investigated in detail with regard to kinematics, kinetics, mass, power and tipover before transitioning into final design. Another common shortcoming in robot design (again speaking from experience!) is to configure a "stick-figure" robot and then learn months later while detailing the design that when the sticks grow into real links that the overall configuration is flawed from the standpoint of clearances, manipulability, mass, deflection or power budget. Since Darwin2K includes considerable detail in each configuration such as gearbox size and mass as well as link wall thickness and stiffness, use of the package avoids unpleasant surprises during the final detailed design phase of a project. I believe that Darwin2K will soon become the basis for advanced robot configuration. It offers the robot designer a rigorous and immensely powerful toolkit with which an increasingly challenging set of robot task design problems can be addressed. Chris has committed to make Darwin2K free, open-source software which should enable its use in advanced governmental and industrial settings as well as educational institutions. As the user base expands, new component libraries and simulation utilities will be added. Task-specific representations xii

11 and task scripting utilities will allow users to leverage upon previous applications of Darwin2K. Ultimately, Darwin2K will be extended to include capability for controller code synthesis as well as sensor configuration which will amplify its usefulness to the robot community. I encourage the reader to utilize this robot configuration methodology and toolkit to create more optimal robot configurations! John Bares Senior Research Scientist The Robotics Institute Carnegie Mellon University May 14, 2000 xiii