Abstract: Interaction with objects and enviornments is at the core of any manipulaton or locomotion behavior, yet, robots mostly try to avoid physical interaction at all costs. This is in stark contrast with humans or animals, that not only constantly interact with their enviornment but also exploit these interactions to their advantage. Reasoning about contact interactions is a computationally daunting task and constitutes one of the main obstacle to robots seamlessly interacting with their enviornments. This talk will present recent efforts towards breaking this complexity, leveraging both optimal control and reinforcement learning algorithms. First, we will show how the structure of physical interactions can be exploited to devise computationally efficient algorithims. Then we will present recent results using Bayesian optimization and reinforcement learning to render such solutions robust to envirornmental uncertainty. We will present experimental results for biped and quadruped robots as well as applications to manipulation.
