Preprint demonstrates energy-focused gait control on a physical quadruped

Researchers report deploying a hierarchical reinforcement-learning controller on a Unitree AlienGo quadruped without additional training on the physical robot. Their preprint separates frequent joint-level control from slower gait adaptation that minimizes the energy needed to travel a given distance. In simulation, the approach reduced that energy measure against single-policy and hierarchical baselines across commanded speeds while maintaining locomotion on flat, rough, and inclined terrain. The learned controller switches from pacing at low speeds to trotting at higher speeds; the comparative energy findings come from simulation.