Letting humans smoothly correct robot hand mistakes mid-task

Vision-Language-Action models accumulate errors during complex hand manipulation tasks, but human correction is itself problematic: when a teleoperator seizes control, the robot's hand configuration snaps discontinuously, causing failures. HandITL blends the human's corrective input with the robot's ongoing policy output so transitions are smooth rather than abrupt. The system also generates higher-quality intervention data for policy refinement — policies retrained on HandITL data outperform those trained on standard teleoperation data by 19% across three long-horizon dexterous tasks involving bimanual coordination and tool use. The work targets robotics researchers and practitioners deploying high-DoF robot hands in contact-rich settings.