Welcome to my home page!
I am a doctoral student at the Media and Machines lab, where I am working with Bill Smart.
UPDATE: Yuval has uploaded the swimmer code for all to use!
I am working on developing local methods for reinforcement learning in continuous domains. I seek ways to train robots to perform complex, articulated motor tasks, without resorting to any ad-hoc dimensionality-reducing approximations or hacks. Most of my work is done in collaboration with Yuval Tassa.
Here are some of our recent demos:
Walkers
A comparison between the initial walker and a walker on stilts:
The composite controller traversing rough terrain:
A comparison between the initial gait and the trained gait for walking uphill:
Swimmers
This is a 14-dimensional system, and all the exhibited behaviors, including the optimal swimming gait, were learned.
A full interaction sequence with a gait learned through receding horizon DDP:
A comparison of the initial and the optimized gait, learned through policy gradient (PG):
A demonstration of a PG swimmer tracing a sigmoid path:
