Abstract:
Enabling robots to execute temporally extended sequences of behaviors is a challenging problem for learned systems, due to the difficulty of learning both high-level task information and low-level control. In this talk, I will discuss three approaches that we have developed to address this problem. Each of these approaches centers on an inductive bias (action primitives, task and motion planning supervision, contact-free motion priors) that enables direct learning of long-horizon behaviors.

Committee:
Ruslan Salakhutdinov
Deepak Pathak
Deva Ramanan
Shikhar Bahl