How to Train a Cross-Embodiment Robot Navigation Policy with AI Agents

NVIDIA's COMPASS workflow adapts an existing robot-navigation policy by training a specialist that corrects its actions for a selected robot and environment. A coding agent coordinates development tasks such as preparing a simulated scene, running initial checks and comparing trained checkpoints. The tutorial keeps human approval at key stages, including accepting the scene and promoting a checkpoint.

The comparison uses matched starting conditions and goals to assess whether the specialist improves on the base policy. At runtime, the trained navigation policy controls motion; the coding agent's role is development and validation. The tutorial ends at checkpoint evaluation, with physical deployment requiring separate checks for the target robot.