Towards Fault-tolerant Quadruped Locomotion with Reinforcement Learning

Modern quadrupedal robots are skilled in navigating through challenging terrains in remote uncontrolled environments with recent advances in reinforcement learning (RL). However, survival in the wild requires not only maneuverability, but also the ability to handle potential critical hardware failures. How to grant such ability to quadrupeds with RL is rarely investigated. In this paper, we propose a novel methodology to enable fault tolerance for RL-based quadruped locomotion controller with joint teacher-student framework for fast zero-shot knowledge transfer that can be deployed to a physical robot without any fine-tuning. With no dedicated reward design for gait guidance, the designed simulation and training strategy can be easily added on top of existing RL-based controllers and generalized to unseen situations. Extensive experiments show that our fault-tolerant controller can efficiently lead a quadruped stably when it faces joint failures during locomotion.