Collects 1M frames (27.8 hours) of egocentric human grasping video and trains a flow-matching model to predict and retarget hand poses to robots. Achieves large zero-shot gains on diverse everyday grasping tasks without requiring robot-specific training data.
