Reinforcement Learning of Serpentine Locomotion for a Snake Robot

The locomotion control of the snake robot is a challenging task that involves many degrees of freedom. Consider the rhythmic behaviour, we introduce the central pattern generator (CPG) model to provide serpentine locomotion, since it can reduce the control dimensionality while remaining continuous and flexible. However, the model parameters are usually chosen empirically, which is time consuming and complicated. Thus, we apply deep deterministic policy gradient (DDPG) algorithm of reinforcement learning (RL) to tune the model and obtain optimal parameters. Consequently, the hyper-redundant snake robot performs graceful serpentine locomotion with an RL controller in simulation. The flexibility of our CPG model and the feasibility of DDPG algorithm are verified according to the experimental results.