Decentralized Reinforcement Learning Multi-Gait Locomotion Control for Quadruped Robots

Multi-gait locomotion is an essential component of various quadruped robot control tasks. However, learning multi-gait locomotion control is challenging due to the considerable differences among gait motions. Meanwhile, these differences make motion imitation difficult to apply to learning multi-gait locomotion control. This work proposes a novel Decomposed Multi-Task Reinforcement Learning (DeMTRL) algorithm that alleviates the difficulty of imitating multiple gaits. The DeMTRL algorithm imitates the decomposed multi-gait leg reference motions with decentralized reinforcement learning. The decentralized learning scheme will exploit the similar leg motion pattern and reduce the difference in imitating different gaits. Additionally, the DeMTRL algorithm can enhance gait stability by considering the morphological symmetry of quadruped robots. Simulation results show that the DeMTRL algorithm can obtain stable real-time multi-gait locomotion and transition.