Learning the robust locomotion on complex terrain combined with model predictive control

Legged robots can adapt to almost any complex terrain because of their unique structure, but it is not easy to design an efficient legged robot controller, which requires a lot of expert knowledge and parameter adjustment. Deep reinforcement learning (DRL) has emerged as a powerful approach for locomotion control of robotic systems, inspired by how animals learn to improve its motor abilities through interaction with the environment. However, it is difficult to synthesize the desired motor behavior through the design of the reward function alone. In this work, We combine traditional control with reinforcement learning controller by adversarial imitation learning. We deployed the trained control on the robot quadruped Gol to achieve a variety of challenging terrain traversal, and we did not use additional visual information, only using the IMU and joint encoder.