Abstract
This paper presents a method to learn stepping motions for fall avoidance by reinforcement learning. In order to overcome the curse of dimensionality associated with the large number of degrees of freedom with a humanoid robot, we consider learning on a reduced dimension state space based on a simplified inverted pendulum model. The proposed method is applied to a humanoid robot in numerical simulations, and simulation results demonstrate the feasibility of the proposed method as a mean to acquire appropriate stepping motions in order to avoid falling due to external perturbations.
