Symmetric Parkour Learning for Robust Quadruped Robot Control on Challenging Terrain

Parkour tasks on various terrains are challenging for quadruped robots, requiring high levels of adaptability and robustness. This paper presents a novel approach to enable quadruped robots to perform parkour tasks on challenging terrain through symmetric reinforcement learning. We propose a modified PPO (Proximal Policy Optimization) that leverages the robot’s symmetry loss to enhance its stability and improve the efficiency of training convergence. This symmetric learning method brings generalization, allowing the robot to be robust even in previously unseen environments. The effectiveness of our approach is demonstrated through results showing both learning convergence efficiency and robust performance on new, challenging terrains when applying the trained model.