Learning CPG-based biped locomotion with a policy gradient method

In this paper, we propose a learning framework for CPG-based biped locomotion with a policy gradient method. We demonstrate that appropriate sensory feedback to adjust the rhythm of the CPG (Central Pattern Generator) can be learned using the proposed method within a few hundred trials in simulations. We investigate linear stability of a periodic orbit of the acquired walking pattern considering its approximated return map. Furthermore, we apply the controllers acquired in numerical simulations to our physical 5-link biped robot in order to empirically evaluate the robustness of walking in the real environment. Experimental results demonstrate that the robot was able to successfully walk using the acquired controllers even in the cases of an environmental change by placing a seesaw-like metal sheet on the ground and a parametric change of the robot dynamics with an additional weight on a shank, which was not modeled in the numerical simulations.