A novel online gait optimization approach for biped robots with point-feet

Abstract

Designing a stable walking gait for biped robots with point-feet is stated as a constrained nonlinear optimization problem which is normally solved by an offline numerical optimization method. On the result of an unknown modeling error or environment change, the designed gait may be ineffective and an online gait improvement is impossible after the optimization. In this paper, we apply Generalized Path Integral Stochastic Optimal Control to closed-loop model of planar biped robots with point-feet which leads to an online Reinforcement Learning algorithm to design the walking gait. We study the ability of the proposed method to adapt the controller of RABBIT, which is a planar biped robot with point-feet, for stable walking. The simulation results show that the method, starting a dynamically unstable initial gait, quickly compensates the modeling error and reaches to a walking with exponential stability and desired features in a new situation which was impossible by the past methods.