RL-Compensated Model Predictive Control for Quadruped Robot Locomotion on Challenging Terrains

Model Predictive Control (MPC) has been widely applied in quadruped robot locomotion. However, the method relies heavily on accuracy of the system model and the feasibility of the pre-planned desired trajectory. Reinforcement learning (RL) improves the performance of the target policy through continuous interaction between the robot and the environment. However, it cannot guarantee the safety of the robot’s actions, and the design of the reward function is a cumbersome process. In this paper, An RL-compensated MPC algorithm framework is proposed. We simplify the quadruped robot into a single rigid body (SRB) dynamic model and train an RL policy to compensate for the linear acceleration, angular acceleration, gait frequency and foothold location of the robot. Comparative experiments against the MPC algorithm in simulation verify that the proposed framework can improve the robot’s locomotion performance on irregular terrains.