Learning and adaptive control for robots

Course's General Objective and Approach

This course presents a wealth of machine learning-based techniques to make the control of robots more flexible and safe when interacting with humans. It introduces a set of control laws that enable reactivity using dynamical systems, a widely used method for solving motion-planning problems in robotics. These control approaches can replan in milliseconds to adapt to new environmental constraints and offer safe and compliant control of forces in contact. The techniques offer theoretical advantages, including convergence to a goal, non-penetration of obstacles, and passivity. The coverage of learning begins with low-level control parameters and progresses to higher-level competencies composed of combinations of skills. 
 
Learning for Adaptive and Reactive Robot Control is designed for graduate students interested in robotics. Techniques covered include learning from demonstration, optimization, and reinforcement learning, and using dynamical systems in learning control laws, trajectory planning, and methods for compliant and force control. 
  •  applications, which range from arm manipulators to whole-body control of humanoid robots
  •  pencil-and-paper and programming exercises
  •  lecture videos, slides, and MATLAB code examples

=========================================================================================

Time and Location of the Course

The course is held on Wednesday morning from

8.15 am - 10.00 am: Lecture - this time may be divided into one hour dedicated to preparation to be done at home and one hour interactive lecture, held simultaneously online and in person.

10.15 am - 12.00 pm: Interactive exercises split into theoretical exercises (pen and paper) and matlab exercises.

The room for the class and exercise is held in INF 213

ROBOT PRACTICE SESSION:

=========================================================================================

Zoom Link

https://epfl.zoom.us/j/62554789973

Only the lecture hours are broacasted on zoom. Exercises and practice sessions are done on site.

=========================================================================================

VIDEOS

Videos recordings of the lectures are posted on the class's video channel on this media channel.  Exercise sessions and matlab practice sessions are not recorded.

=========================================================================================

Course Material

The Lecture is based on the book:
Billard, Aude, Sina Mirrazavi, and Nadia Figueroa. Learning for Adaptive and Reactive Robot Control: A Dynamical Systems Approach. MIT Press, 2022.

PURCHASE of Hardcopies: since EPFL bookstore closed, a few copies of the book can be purchased directly from LASA laboratory secretary. Please, contact the teaching assistants of the class. Alternatively, you can purchase hardcopies or Kindle versions from Amazon.

ELECTRONIC VERSION: An electronic copy is available through the EPFL library via the BEAST Catalog or directly on the provider at this LINK. You are required to login via VPN or be on campus to access the book.

RENTAL:  the book can be rented as e-textbook rentals through  https://mitpress.ublish.com/

=============================================================================================

ADDITIONAL READINGS

• Handbook of Robotics – Edited by B. Siciliano and O. Khatib - Springer-Verlag – available for free on-line - This book offers comprehensive overview of basics of robot control  Most relevant chapters for this class are: “Kinematics,” “Motion Planning,” “Motion Control” and “Force Control”.

• Applied Nonlinear Control - J-J. Slotine and W. Li - Prentice Hall – available for free on-line -- This book offers comprehensive introduction to dynamical systems for control. Most relevant chapters are "Phase Plane Analysis” and “Fundamental of Lyapunov’s Theory.”
  

=========================================================================================

SOFTWARE

The course uses Matlab. You must bring your own laptop and should use a version of Matlab 2019 and higher.

Instructions to access the EPFL virtual machine are available HERE.

========================================================================================

Instructors
Prof. Aude Billard  
LASA Laboratory
Swiss Federal Institute of Technology - EPFL
CH-1015 Lausanne, Switzerland

email: aude.billard@epfl.ch

Office Hours: Thursday, 13:00 to 14:00, by appointment. (room ME.A3.393) 
Tel: +41 (21) 693.54.64
fax: +41 (21) 693.78.50

=================================================================

Teaching Assistants

• Erfan Etesami erfan.etesami@epfl.ch
• Yang Liu, yang.liuu@epfl.ch
  

• Elise Jeandupeux elise.jeandupeux@epfl.ch 

==================================================================

GRADING / EXAM

Grade is based 100% on an oral exam that will be held at the end of the semester during the regular exam session. A doodle poll will be sent to choose the time of the exam. The dates of the exam will be announce by the academic services during the spring semester.

===================================================================

BOOK CHAPTER TO READ: Chapter 1

The lecture is followed by two hours of matlab exercises.

BOOK CHAPTER TO READ: Chapter 2

It is complemented with 1 hour of pen and paper exercises and 1 hour of matlab implementation of some of these exercises.

BOOK CHAPTER TO READ: Annexes A

The lecture is complemented with 1 hour of pen and paper exercises, followed by 1 hour of matlab exercises during which you will be able to test the sensitivity of these algorithms to choice of hyperparameters.

BOOK CHAPTER TO READ: Chapter 3 - Sections 3.1, 3.2, 3.3 and 3.4, Annexes B3.1-3.3

Time: 8:15-12:00

Objective: This is a 4-hour practice session in matlab. You will systematically assess the power of the two algorithms seen in class for learning stable control laws with dynamical system for a 7-degrees of freedom robot arm.

Objective: This lecture introduces the notion of local modulation. It will show how one can modulate an initial simple linear dynamical system to make it locally non-linear. It will show how such modulation can be learned from data.

The lecture is accompanied with pen and paper exercises and matlab exercises to familiarize you with the theoretical construct of the modulation and its implementation for learning from handdrawn data.

BOOK CHAPTER TO READ: Chapter 8

The lecture is followed by a 45 minutes pen and paper exercise session and another 45 minutes session on matlab to familiarize you with the theoretical construct of the obstacle avoidance modulation and its implementation.

BOOK CHAPTER TO READ: Chapter 9

Objective: 

This lecture introduces the concept of impedance control and its role in enabling robots to make safe contact with their environment. It then presents ways by which the impedance can be learned through human demonstrations. Finally, it offers one solution to combine dynamical systems control and impedance control to enable safe control of robots.

The lecture is followed by a 45 minutes pen and paper exercise session and another 45 minutes session on matlab to familiarize you with the theoretical construct of the obstacle avoidance modulation and its implementation.

Book Correspondence: Chapter 10, sections 10.1, 10.2, 10.3 and 10.4

Objective: 

This lecture presents method to perform torque control of robots through DS. It also shows how the modulation in force can be taught to the robot either offline or live.

The lecture is followed by a 45 minutes pen and paper exercise session and another 45 minutes session on matlab to familiarize you with the theoretical construct of the obstacle avoidance modulation and its implementation.

Book Correspondence: Chapter 10, section 10.4 & Chapter 11

No class.

Time: 8:15-12:00

Objective: 

This is a 4-hour practice session in matlab. You will explore the DS modulation method for obstacle avoidance in 3D using the simulated 7-degrees of freedom robot arm.

The lecture will start with a brief a overview of the class material and present the exam format.

It will then be followed by a presentation of several extensions to the method we have seen in class to use DS for enabling coupling across dynamical systems and their multiple applications for  coordinated control of arm and hand, and for coordinated control of multi-arm systems, with examples for catching objects in flight, or cutting tissues. This additional material will not be part of the material examined during the exam.

Book Correspondence: Chapters 6 and 7

Practice Session :

This practice session will allow you to test the different algorithms seen in class when controlling real robots. It will replace the exercise sessions until the end of the semester.

Time: 8:15-13:00

Objective: 

This practice session will allow you to test the different algorithms seen in class when controlling real robots. It will replace the exercise sessions until the end of the semester.

Time Practical: 8:15-12:00

This practice session will allow you to test the different algorithms seen in class when controlling real robots. It will replace the exercise sessions until the end of the semester.

The exam dates and location are decided by the EPFL Academic Services, not the teacher. They are usually announced in  May.