Sim-To-Real: Learning Energy-Efficient Slithering Gaits for a Snake-Like Robot

To resemble the body flexibility of biological snakes, snake-like robots are designed as a chain of body modules, which gives them many degrees of freedom on the one hand and leads to a challenging task to control them on the other hand. Compared with conventional model-based control methods, reinforcement learning based methods provide promising solutions to design agile and energy-efficient gaits for snake-like robots, since reinforcement learning based methods can fully exploit the hyper-redundant bodies of the robots. However, reinforcement learning based methods for snake-like robots have rarely been investigated even in simulations, let alone been deployed on real-world snake-like robots. In this work, we introduce a novel approach for designing energy- efficient gaits for a snake-like robot, which first learns a policy using a reinforcement learning algorithm in simulation and then transfers it to the real-world testing, thereby leveraging fast and economical gait generation process. We evaluate our reinforcement learning based approach in both simulations and real-world experiments to demonstrate that it can generate substantially more energy-efficient gaits than those generated by conventional model-based controllers.