Terrain-guided Symmetric Locomotion Generation for Quadrupedal Robots via Reinforcement Learning

Most living organisms in the natural world exhibit inherent symmetry during their locomotion processes. This biological symmetry has inspired this paper to design symmetry reward functions for non-trajectory-based training process of Reinforcement Learning. As a result, the training time to obtain robust policies can be greatly reduced. Besides, this paper also designs a continuously changing terrain randomization routine for the training process, the effect of which indicates that legged locomotion with its intermittent ground contact is the natural choice on uneven rough terrains. The experimental results from stable walking and running behaviors of a quadrupedal robot demonstrate the effectiveness of the obtained policies. The robot can also reject disturbances like human kicks reasonably well during locomotion. Additionally, through extra velocity randomization during the policy training process, a single policy network can be obtained to help the robot track different target velocities actively.