Learning Arm Motion Strategies for Balance Recovery of Humanoid Robots

Abstract

Humans are able to robustly maintain balance in the presence of disturbances by combining a variety of control strategies using posture adjustments and limb motions. Such responses can be applied to balance control in two-armed bipedal robots. We present an upper-body control strategy for improving balance in a humanoid robot. Our method improves on lower-body balance techniques by introducing an arm rotation strategy (ARS). The ARS uses Q-learning to map sensed state to the appropriate arm control torques. We demonstrate successful balance in a physically-simulated humanoid robot, in response to perturbations that overwhelm lower-body balance strategies alone.