COMPASS: Cross-embOdiment Mobility Policy via ResiduAl RL and Skill Synthesis

Technical Analysis “COMPASS” is a unified framework that enables scalable cross- embodiment mobility using expert demonstrations from only a single embodiment. High-level Analogy: Imagine you want to teach a group of aspiring drivers to navigate various terrains, but they each have a different type of vehicle (a small car, a tall truck, a nimble motorcycle, a heavy bus). General Driving School (Imitation Learning): First, you train everyone in a standard car (like a common sedan). You teach them the basic rules of the road, how to steer, brake, and accelerate, and how to avoid obstacles using clear, easy-to-understand demonstrations. This gives everyone a strong 'base' understanding of driving. Specialized Coaching (Residual Reinforcement Learning): Now, each driver takes their own unique vehicle. The 'truck' driver gets specific coaching on how to handle its height and blind spots, the 'motorcycle' driver learns to lean into turns, and the 'bus' driver learns about its long turning radius. Instead of re-teaching them everything from scratch, the coach just gives them small, corrective tips to adapt their general driving skills to their specific vehicle's quirks. They already know how to drive; they just need tweaks. The Super Driver Handbook (Policy Distillation): Finally, you gather all the lessons learned from each specialized coach. You write a 'Super Driver Handbook' that combines all this knowledge. This handbook isn't just for one vehicle; it has sections for 'if…