A Reinforcement Learning Path Following Strategy for Snake Robots Based on Transferable Constrained-Residual Gait Generator

Due to the high-redundant multilink body configuration and complex ground contact model, accurate and flexible path following is difficult to accomplish for planar snake robots, which has always attracted the attention of many scholars. Many traditional path following controllers can successfully drive a snake robot to follow comparatively simple paths. Yet, for more complex paths, such as paths with discontinuous curvature, the accuracy and generalizability are always unsatisfactory. Therefore, in this article, a novel reinforcement learning (RL) strategy based on transferable constrained-residual gait generator (CRGG) is proposed for path following of snake robots, which ensures accurate, efficient, and agile following for any target paths. Specifically, first, to enhance the generalizability of the policy in various path following tasks, the training for specific path following is converted into the training for random points following, which gives a snake robot good generalizability for any target paths represented by a set of target points. Subsequently, the transferable RL algorithm CRGG is proposed to regulate the constrained motion gait while concurrently compensating for it in a residual manner, in which the following performance is guaranteed, especially for the complex paths with discontinuous curvature. Simulation and sufficient experimental results are provided to illustrate the superior performance of the proposed RL controller in terms of path following ability, generalizability, and robustness.