Abstract: In robotics, it is essential to be able to plan efficiently in high-dimensional continuous state-action spaces for long horizons. For such complex planning problems, unguided uniform sampling of actions until a path to a goal is found is hopelessly inefficient, and gradient-based approaches often fall short when the optimization manifold of a given problem is not smooth. In this talk, I will introduce two of our recent approaches for using past planning experience to learn to predict constraints for guiding a planner. In the first part, I will introduce a technique that uses generative adversarial networks and importance sampling estimation to learn an action distribution that restricts the search space to a promising region, when an input is represented with a vector. In the second part, I will discuss the limitation of a vector representation in complex robot planning settings, and propose a more suitable representation for guiding a search, called a score-space representation. With this representation, we can predict a constraint on the search space by optimizing a black-box function. We empirically show that a planner is able to find a solution more efficiently using these approaches on a various robot task and motion planning problems.
Bio: Beomjoon Kim is a PhD student at MIT CSAIL under the supervision of Leslie Pack Kaelbling and Tomas Lozano-Perez. His recent research focuses on developing machine learning algorithms for complex robot planning problems, in which problems involve reasoning about both discrete, logical structures and continuous, geometric structures of the world. In the past, he has worked on robot learning from demonstrations and reinforcement learning. He received his MS.c from McGill University under the supervision of Joelle Pineau, and received BMath from University of Waterloo.