Research / arXiv / Oct 31, 2024
EgoMimic: Scaling Imitation Learning via Egocentric Video
EgoMimic frames human egocentric video as scalable demonstration data for robot manipulation and cross-embodiment policy learning.
EgoMimic investigates how human first-person manipulation footage can expand robot learning beyond demonstrations collected only on the target robot.
The approach makes camera viewpoint, hand visibility, and object interaction quality central to the usefulness of each recorded task sequence.
The original paper provides the training design, experimental comparisons, and the authors' discussion of transfer limitations.
egocentric visionimitation learningrobot manipulationhuman demonstration
Open original source