A simple rebalance strategy for omnidirectional humanoids walking by learning foot positioning

Abstract

On solving the rebalance problem of the trajectory-based humanoids walking approaches, a simple foot positioning compensator is proposed to modify the foot positioning online based on the estimated robot state using onboard sensors. To make the compensator coincident with the dynamics of a full-body humanoid robot, the foot positioning policy is learnt through a policy gradient reinforcement learning approach. Experiments on both simulated and real full-body humanoid robots validate the good performance of the proposed method not only in forward walking but also in omnidirectional walking.