Reinforcement learning based CPG controller for a soft gastropod robot and its gaits comparison study

Bio-inspired robots, such as soft gastropod-like robots, have been developed to achieve environmental compatibility locomotion as their natural counterparts. However, it is difficult to determine the optimal locomotion gaits of these robots, as existing gaits in the biological animals are not easy to evaluate and compare. To investigate this problem, we propose a framework that employing deep reinforcement learning (deep RL) for high-level motion planning and central pattern generators (CPGs) for low-level motion control. Instead of controlling the robot with manually designed rules, the framework enables robots to select and control the locomotion gaits by themselves. This framework is verified with a soft gastropod robot. To train the RL agent, a virtual robotic locomotion environment is built based on the robot kinematics. Simulations and physical experiments show that the control framework can well help the robot to switch and control its three gaits to reach the target,