Learning-based velocity-controlled bipedal locomotion with heuristic footstep guidance: revisiting the role of model-based priors

A heuristic approach to step planning for bipedal locomotion learning is presented, extending the standard control scheme based on tracking the robot’s desired torso velocity. This enhancement enables precise interaction with the environment, particularly in overcoming obstacles such as steps and gaps, and in accurately approaching target locations. Unlike methods based on full or simplified dynamics, the proposed solution does not require complex footstep planners or rely on a priori models. Step planning is performed using simple Raibert-style heuristics and is implemented through feedback between the planner and the learned policy. The approach is compared with two common model-based step planning methods: one based on the Linear Inverted Pendulum Model (LIPM) and the other on fullbody dynamics via Hybrid Zero Dynamics (HZD). Experiments demonstrate comparable or superior speed-tracking accuracy (up to 80%) and significantly greater stability on rough terrain (up to 50% improvement), while maintaining similar energy efficiency. These results suggest that incorporating complex model-based components into the training architecture may be unnecessary for achieving robust bipedal locomotion.