Abstract:
In this talk, I focus on efficient online learning for solving real-world physics puzzles. I discuss challenges associated with learning in this domain and how those challenges inform certain design decisions. In particular, learning from scratch in the real world would be difficult. I present a practical mixture of experts framework for learning strategies to solve physics puzzles in simulation and efficiently refine those strategies in the real world with online learning. Mixture of experts policies are capable of representing multiple possible strategies for the same or similar tasks and, thus, promote efficient learning by enabling exploration of multiple promising regions of the action space. I also describe the robotic system I developed to evaluate learning algorithms on real-world physics puzzles. With the proposed framework, online learning more than doubles the performance on the robot.

Committee:
Chris Atkeson
Oliver Kroemer
David Held
Leo Keselman