Learning to Walk by Actuating a Passive Dynamic Walker

Russell L. Tedrake
MIT Brain & Cognitive Sciences Department

March 25, 2005
11 a.m. - 12 noon
Fuller Labs 320

Abstract

Humanoid robots are an exciting new form of interactive media, but designing control systems for walking robots has turned out to be an extremely difficult problem. Bipedal robots today can walk across a flat factory floor, and even up stairs, but they cannot compete with humans in terms of speed, efficiency, or robustness. An essential component of human walking which allows us to achieve this performance is our ability to learn and adapt to the terrain as we walk. In order to develop robots that can optimize their control policies in real-time, we use an old trick from the walking literature - "passive dynamic walking". Passive dynamic walkers are a class of robots which are capable of walking down a small decline without using any motors or control. By adding a small number of actuators to a machine that is already capable of passive walking down a ramp, we are able to optimize a feedback control policy for walking on flat terrain in real-time; the robot "learns to walk" in approximately 20 minutes without any initial control knowledge, and continually adapts to the terrain as it walks. More fundamentally, by formulating the goal of walking as an optimization problem, we can develop novel controllers which generate robust and efficient locomotion for a variety of machines over a variety of terrain.

Biography

Russ Tedrake is a Postdoctoral Associate with the MIT Brain & Cognitive Sciences Department. He received his Ph.D. from the MIT Computer Science & Artificial Intelligence Laboratory in September, 2004, working with Sebastian Seung, the MIT Leg Laboratory, Emilio Bizzi, and Jean-Jacques Slotine. His primary research focus is using machine learning and optimal control to improve the state-of-the-art in walking robots.

Host

Michael A. Gennert
Refreshments will be served.

Maintained by webmaster@wpi.edu
Last modified: Sep 27, 2006, 16:05 EDT