A New Approach to Control of Legged Robots

Designing dynamically stable controllers for a robot with 2r legs is challenging due to its complex hybrid dynamics (r 1). This paper proposes a technique to decompose the robot into r biped robots, where the influence of other robot parts on each biped can be modeled as external forces. This approach allows existing research on biped control to be applied to the quadruped robot. Time-invariant controllers, which typically ensure walking stability for planar point-footed bipeds, are selected for this purpose. For clarity, we focus on a planar point-footed quadruped for decomposition. We extend a recent reinforcement learning method to optimize these controller parameters for walking on slopes or under specific forces, while accounting for significant modeling errors in the quadruped. Simulation results demonstrate that our method achieves stable walking with the desired features and effectively compensates for modeling errors.