Learning to Adapt: Bio-Inspired Gait Strategies for Versatile Quadruped Locomotion

Deep reinforcement learning (DRL) has revolutionised quadruped robot locomotion, but existing control frameworks struggle to generalise beyond their training-induced observational scope, resulting in limited adaptability. In contrast, animals achieve exceptional adaptability through gait transition strategies, diverse gait utilisation, and seamless adjustment to immediate environmental demands. Inspired by these capabilities, we present a novel DRL framework that incorporates key attributes of animal locomotion: gait transition strategies, pseudo gait procedural memory, and adaptive motion adjustments. This approach enables our framework to achieve unparalleled adaptability, demonstrated through blind zero-shot deployment on complex terrains and recovery from critically unstable states. Our findings offer valuable insights into the biomechanics of animal locomotion, paving the way for robust, adaptable robotic systems.