Discover skills for integrating locomotion with downstream tasks

Determine methods to automatically discover and learn the set of locomotion skills necessary to integrate legged locomotion with downstream tasks such as loco‑manipulation, so that long‑horizon feasibility and task completion can be achieved in real‑world settings.

Background

The survey highlights that while legged locomotion controllers have matured, integrating locomotion with downstream tasks (for example, loco‑manipulation or parkour) requires a repertoire of skills beyond basic walking and recovery. Training end‑to‑end policies for such tasks is complicated by long‑horizon objectives, sparse rewards, and complex contact dynamics.

The authors note that it is not yet clear how robots should discover the skills needed for downstream tasks, and that principled approaches to learning and composing these skills are still missing.