Egocentric dataExpert
Ego4WAM: What Matters When Scaling Egocentric Human Data for Robot Learning?
Preprint. The paper presents Ego4WAM, a controlled study of how to scale and stage egocentric human data for bimanual robot manipulation under a world-action model. Its headline result is that video-only world-model pre-training on 15K hours of egocentric video plus only about 0.1K hours of robot data reaches a 20.29 average RoboDojo score and 14.35% success rate, roughly matching or exceeding several models trained with far more robot data. The work matters because it identifies which data properties—human-robot alignment, duration, task diversity, and supervision—actually improve closed-loop robot generalization.