CSAIL Event Calendar: Previous Series
Learning to Walk by Exploiting Passive Dynamics (with live robot demo)
Speaker: Russ Tedrake , The Seung Lab
Bipedal walking robots today can walk across a flat factory floor, and even up stairs, but even our best bipedal control algorithms leave much to be desired in terms of stability on rough terrain and in terms of energy efficiency. Many researchers have tried to improve these control systems using reinforcement learning, but have struggled with large continuous state and action spaces and with the minimal number of learning trials that you can afford to run on a real robot. I will present a simple bipedal walking robot that we have designed, using the concept of passive dynamic walking, which simplifies the problem to the point where it is possible to train an actor-critic style reinforcement learning algorithm completely on the robot. The learned controller is quantitatively more stable than any hand-designed controllers that we were able to derive. I will bring the robot to the talk for a live demonstration.