A Motion Planning and Control Method of Quadruped Robot Based on Deep Reinforcement Learning

摘要

Motion planning and control issue of quadruped robots is a challenging topic because it requires a lot of professional knowledge and tedious manual design based on traditional control methods. This paper proposes a novel method of training quadruped robots to generate locomotion gaits by combining the deep reinforcement learning. The control problem of quadruped robots is modeled as a Markov Decision Process, and a universal quadruped robot motion control reward mechanism is developed. A variety of gaits for the quadruped robot is generated by adding a gait reference frame. Two deep reinforcement learning algorithms are used for training and comparison. The desired gait locomotion control strategy is generated through training so that the robot can realize stable walking with the desired gait. Finally, the proposed method is tested and evaluated in a simulation environment using the A1 quadruped robot model.