Deep Reinforcement Learning for Snake Robot Locomotion

The design of gaits for underactuated robots is often unintuitive, with many results derived from either trial and error or simplification of system structure. Recent advances in deep reinforcement learning have yielded results for systems continuous in either states or actions, which may extend to a variety of locomoting robots. In this work we employ reinforcement learning to derive efficient and novel gaits for both terrestrial and aquatic multi-link snake robots. Although such systems operate in different environments, we show that their shared geometric structure allows us to utilize the same learning techniques in both cases to find gaits without any human input. We present results learned and rolled out in simulation, and we describe preliminary efforts to implement the entire learning process on a physical system.