A stochastic policy gradient based adaptive control for biped walking

This paper investigates the walking control problem of a planar compass-like biped robot. Suppose that the robot model is not known. A stochastic policy gradient based adaptive control law is then proposed for the purpose of realizing a stable walking with a desired step length. As we show in this paper, with only a constant control input, the robot under consideration has a stable one-step limit cycle when the constant control input is small (equivalently, the step length is small), but when the constant control input becomes larger to make a larger step length, the robot walking trajectory diverges away from the one-step limit cycle and converges to an undesirable two-step limit cycle. Our proposed stochastic policy gradient based adaptive control can overcome the deficiency of using the constant control and realize stable walking over the one-step limit cycle with any desired step length. The validity of the proposed control method is verified by simulation results. Moreover, from the simulation results, the basin of attraction is also enlarged by comparing with other event-based control for biped walking.