Computing Reviews
Today's Issue Hot Topics Search Browse Recommended My Account Log In
Review Help
Search
Reinforcement learning of bimanual robot skills
Colomé A., Torras C., Springer International Publishing, New York, NY, 2019. 182 pp. Type: Book (978-3-030263-25-6)
Date Reviewed: Feb 16 2021

This book is an extension of Colomé’s doctoral dissertation, which was directed by Torras. The foreword credits it as “a finalist for the 2018 Georges Giralt PhD Award for the best doctoral thesis in Europe,” though actually only robotics theses were considered. In 2018, there were 27 submissions, with two winners chosen from just five finalists. In other words: it was still a stellar contribution.

This book is a close copy of the dissertation, with minor extensions of discussion and analysis, as well as the additions of chapter 8 and Appendix A (described below). In fact, readers can freely consult the thesis online (https://upcommons.upc.edu/handle/2117/119092) before buying the book. It appears that a book development project was launched with the thesis, for its typesetting and art largely went to press. Indeed, several mentions of “the authors” in the book can also be found in the thesis.

The book provides explanations, examples, mathematical analysis, and experiments designed to champion novel proposals for improved closed-loop inverse kinematics (CLIK) algorithms, compliant dynamics (requirements that assure safety when robots operate near humans and objects), and reinforcement learning. Dimensionality reduction (DR) is an especially essential and effective contribution. Appendix B lists videos of experiments and other materials available online.

The focus is on redundant robots, whose degrees of freedom are strictly dominated by the number of task-space dimensions. This provides flexibility, but offers infinitely many IK solutions for a desired pose. Coping with this computational difficulty is a principal issue of Part 1 (chapters 2 to 4). Part 2 (chapters 5 to 8) turns to two models of movement primitives, dynamic (DMP) and probabilistic (ProMP), to encode robot motion policies. Chapter 7 leads to DR for both models.

Chapter 8 was added after the thesis was completed to show contextual abilities for ProMP, including obstacle avoidance and the accommodation of several robot actions that share a context feature. New experimental examples include drawing an alphabet and feeding a mannequin. The added Appendix A provides the mathematics to perform DR, not in the robot-joint task space (done in chapter 7) but in the parameter space for both DMP and ProMP. Final examples of what was added to the thesis are Sections 5.2.2.1 and 5.2.2.2; there are probably more.

Readers familiar with robotic manipulation will appreciate this work the most, as it is situated in the most recent journal articles and conference papers. But even a neophyte such as I (a casual acquaintance with Ken Goldberg and Sandy Pentland; a midnight visit to Cynthia Breazeal’s lab; a shallow but deepening appreciation for “Do you love me?”) can gain perspective on the problems faced, the advances made, the methods introduced, and the related problems that remain. Each chapter, no matter how technical, leads off with a lucid introduction. There are particularly fine summaries at the end of each chapter and in chapter 9.

Unfortunately, numerous typographical errors carry over from the original thesis. While none I saw are ruinous, they are annoying. Some examples:

  • Words have been added to the book’s outline, but a useful figure (a thesis overview) has been omitted. Figure 1.1 remains with a confusing double-headed arrow. At least the weak title of “Thesis block diagram” has been improved to “Block diagram of our approach to learning robot motion skills.”
  • At the bottom of page 13, J* should be bold roman, not light italic. This is repeated six lines later, and then the * is dropped from J in Formula (2.5).
  • Figures 3.6 and 3.8 appear in other literature and deserve better settings. The latter has no alphas, although colors do distinguish the various angles. References in Figure 4.6 must point somewhere else.
  • Formula (7.12) needs to be squared with the second next sentence.

Particular to the book:

  • While the book fixes a typographical error on page 110--“eliminatint” in the thesis was replaced with “eliminating”--it fails in the first line of the first chapter when it creates “neuro physiologist.”
  • A math instance is the dropping of an inappropriate P on top of a dotted y in Formula (5.14).
Reviewer:  Benjamin Wells Review #: CR147189 (2106-0139)
Bookmark and Share
  Reviewer Selected
Featured Reviewer
 
 
Learning (I.2.6 )
 
 
Real-Time And Embedded Systems (C.3 ... )
 
 
Robotics (I.2.9 )
 
Would you recommend this review?
yes
no
Other reviews under "Learning": Date
Learning in parallel networks: simulating learning in a probabilistic system
Hinton G. (ed) BYTE 10(4): 265-273, 1985. Type: Article
Nov 1 1985
Macro-operators: a weak method for learning
Korf R. Artificial Intelligence 26(1): 35-77, 1985. Type: Article
Feb 1 1986
Inferring (mal) rules from pupils’ protocols
Sleeman D.  Progress in artificial intelligence (, Orsay, France,391985. Type: Proceedings
Dec 1 1985
more...

E-Mail This Printer-Friendly
Send Your Comments
Contact Us
Reproduction in whole or in part without permission is prohibited.   Copyright 1999-2024 ThinkLoud®
Terms of Use
| Privacy Policy