Learning Spring Mass Locomotion: Guiding Policies With a Reduced-Order Model

In this letter, we describe an approach to achieve dynamic legged locomotion on physical robots which combines existing methods for control with reinforcement learning. Specifically, our goal is a hierarchy in highest-level behaviors are planned through reduced-order models, the fundamental physics of locomotion, and lower level controllers utilize learned policy that can bridge gap between idealized, simple model complex, full order robot. The high-level planner use environment be task specific, while low-level controller execute wide range motions so it applies many different tasks. walking show from models used as command primary training signal policies. resulting policies do not attempt naively track motion (as traditional trajectory tracking would) but instead balance immediate long term stability. demonstrated human scale, unconstrained, untethered bipedal robot at speeds up 1.2 m/s. This letter builds foundation generic, applied