Can one robot arm teach two-arm tasks?

Learning bimanual manipulation is hard because two-arm robots are rare. MonoDuo solves this by having one robot arm perform half a task while a human does the other, then swapping roles. The system synthesizes realistic demonstrations using hand pose estimation and image inpainting, then trains bimanual policies on these augmented data. On five real tasks—box lifting, cloth folding, jacket zipping—the approach reaches 70% success zero-shot and 65–70% improvement with just 25 real target demonstrations.