Information-Geometric Policy Search for Learning Versatile, Reusable Skills

In this talk, I will present our work on information-geometric policy search methods for learning complex motor skills. Our algorithms use information-geometric insights to exploit curvature and path information in order to perform efficient local search at the level of single elemental motions, also called movement primitives. Simultaneously to local search, the algorithms search on a global level by selecting between distinct solutions, allowing us to represent a versatile solution space with high quality solutions. Our algorithms can be used to efficiently learn motor skills, generalize these motions to different situations, learn reactive skills that can react to perturbations and select and learn when to switch between these motions. I will also briefly show how to extend our algorithms to learn from preference-based feedback instead of a numeric reward signal, enabling a human expert to guide the learning agent without the need for manual reward tuning. While I will use dynamic motor games, such as table tennis, as motivation throughout my talk, I will also shortly present how to apply similar methods for robot grasping and manipulation tasks.