Berkeley AI Research

“The authors build an end-to-end pipeline that ingests raw monocular RGB video (YouTube clips, egocentric footage, AI-generated video) and outputs physically executable trajectories for a 22-DoF dexterous robot hand.”

“In total, our pipeline produced 500 high-quality, human-verified dexterous manipulation trajectories across internet (53%), egocentric (31%), and generated (16%) videos.”