A Residual Meta-Reinforcement Learning Method for Training Fault-Tolerant Policies for Quadruped Robots

Motor locking is a common issue in quadruped robots that can have serious consequences if the robot continues executing its original commands. However, the static stability of the quadruped allows for the flexibility to adjust the robot's control policy so that it can maintain movement along a predetermined trajectory. In this paper, we introduce a residual meta reinforcement learning method comprising a trajectory generator and a meta-reinforcement learning corrector. The trajectory generator generates a reference joint position, while the corrector utilizes contextual reasoning to determine the appropriate action in the event of a motor locking. This action is employed to rectify the reference joint position, resulting in a fault-tolerant control strategy for the robot. We conducted comprehensive simulation experiments to validate our proposed algorithm, which demonstrates that the robot can still follow the predefined trajectory, even in the presence of a motor locking. Moreover, our proposed approach outperforms all baseline algorithms.