Online learning of a full body push recovery controller for omnidirectional walking

Bipedal humanoid robots are inherently unstable to external perturbations, especially when they are walking on uneven terrain in the presence of unforeseen collisions. In this paper, we present a push recovery controller for position-controlled humanoid robots which is tightly integrated with an omnidirectional walk controller. The high level push recovery controller learns to integrate three biomechanically motivated push recovery strategies with a zero moment point based omnidirectional walk controller. Reinforcement learning is used to map the robot walking state, consisting of foot configuration and onboard sensory information, to the best combination of the three biomechanical responses needed to reject external perturbations. Experimental results show how this online method can stabilize an inexpensive, commercially- available DARwin-OP small humanoid robot.