Efficient and safe real‑world learning for quadruped locomotion

Develop sample‑efficient and safe reinforcement learning procedures for real‑world training and adaptation of quadruped locomotion policies on hardware, minimizing human intervention while ensuring reliability and safety.

Background

Zero‑shot sim‑to‑real transfer has yielded robust quadruped controllers, but fine‑tuning and learning directly in the real world remain challenging due to safety risks, limited recovery capabilities, and the need for automatic resets.

The authors identify efficient and safe real‑world learning as a core unsolved issue in mature quadruped locomotion, necessary for broader deployment and adaptation to novel conditions without relying on high‑fidelity simulation.