Autonomous bipedal walking pace supervision under perturbations

This paper presented a method of bipedal walking pace supervision by the adjustment of stride-frequency and step-length simultaneously. A reinforcement learning algorithm is designed to learn the walking stride-frequency; A transition plan aims to adjust the step-length or update motion phases according to the dynamic feedback; A momentum based estimation gives another layer of stride-frequency adjustment when the learning agent has not gained enough experiences. Simulation experiments showed this learning based motion supervision is effective for maintaining stable walking under perturbations with a balanced performance of energy consumption and robustness.