Turning ordinary human videos into skill-labelled training data for a dexterous robot handIntermediate
DexStitch: Learning and Composing Dexterous Skills from Human Videos via Contact-Aware Robotic Inpainting
A Korean preprint from Seoul National University, Hyundai Motor Company's Robotics Lab and DGIST reaches 77.5% success on four-skill household sequences with a dexterous hand, where the UniSkill baseline trained on the same data scores 0%. The trick is to label the motion between skills as a skill of its own and to learn it from about 240 human videos plus only 32 short teleoperated robot episodes.