Transferring task structure from human videos to robot world-action modelsExpert
Cross-Embodiment Causal Transfer from Human Videos to Robot Manipulation
Human videos are cheap and plentiful, but robots cannot copy human hand motions directly. This preprint from the Chinese Academy of Sciences and Meituan extracts only the task logic from such videos and feeds it to an existing robot model, adding about one point in near-saturated simulators and 8 points on a real dual-arm robot.