SEED-UMI: Sharing Exoskeletons Across Human and Robot for Direct Dexterous Demonstration

SEED-UMI tackles the cross-embodiment gap in dexterous manipulation by having both the human operator and the robotic hand wear the exact same exoskeleton hardware. This shared physical interface aligns encoder measurements and ensures wrist-mounted cameras observe the identical mechanical structure during both demonstrations and autonomous rollouts, removing the need for open-loop retargeting or visual inpainting. Across five contact-rich tasks, the method achieves 3.0x faster data collection than exoskeleton teleoperation and a 70.0% average rollout success rate.