PPO-CPG Controller for Quadruped Locomotion

The central pattern generator (CPG) can produce stable rhythmic signals for the locomotion of quadruped robots. However, the single CPG-based approach still requires further command tracking and movement stability enhancement due to its inherent lack of sensory feedback signals. In this paper, we proposed a PPO-CPG control architecture, which integrated deep reinforcement learning (DRL) techniques to enable the agent to modify the input parameters of the CPG network directly, thereby adjusting both the amplitude and frequency of gait output signals for achieving more stable and natural locomotion patterns. We conducted training within the Isaac Gym simulation environment and evaluated its performance on rough terrain conditions. The results indicated that our method reduces speed tracking error by approximately 55.45% compared to a single CPG-based method, demonstrating a significant improvement.