Stable and efficient learning of long‑horizon navigation with low‑level control in the loop

Develop reinforcement learning methods that can stably and efficiently learn long‑horizon navigation policies while incorporating low‑level locomotion control during training, addressing sparse rewards, contact dynamics, and training instability.

Background

Integrating navigation and locomotion has enabled agile legged and aerial navigation, but training becomes difficult due to long horizons, sparse rewards, and complex low‑level dynamics.

The authors point out that despite promising results, reliable and scalable learning procedures for long‑horizon navigation when low‑level control is in the loop are still lacking.