State Representation Learning from Demonstration

Astrid Merckling, Alexandre Coninx, Loic Cressot, Stéphane Doncieux, Nicolas Perrin

Abstract: Robots could learn their own state and universe representation from perception, experience, and observations without supervision. This desirable goal is the main focus of our field of interest, State Representation Learning (SRL). Indeed, a compact representation of such a state is beneficial to help robots grasp onto their environment for interacting. The properties of this representation have a strong impact on the adaptive capability of the agent. Our approach deals with imitation learning from demonstration towards a shared representation across multiple tasks in the same environment. Our imitation learning strategy relies on a multi-head neural network starting from a shared state representation feeding a task-specific agent. As expected, generalization demands tasks diversity during training for better transfer learning effects. Our experimental setup proves favorable comparison with other SRL strategies and shows more efficient end-to-end Reinforcement Learning (RL) in our case than with independently learned tasks.