EgoMimic: Scaling Imitation Learning via Egocentric Video

EgoMimic is a full-stack framework for scaling robot manipulation with human embodiment data. It records egocentric human video with 3D hand tracking and treats those demonstrations as a major training source alongside robot data.

The system combines an ergonomic capture setup, a low-cost bimanual manipulator, cross-domain data alignment, and an imitation-learning architecture that co-trains on human and robot demonstrations.

The research focuses on long-horizon single-arm and bimanual tasks and reports improvements over the evaluated imitation-learning baselines, including generalization to new scenes.

Its scaling study reports that an additional hour of human hand data produced greater value than an additional hour of robot data under the paper's experimental setting. That result highlights the importance of diverse, well-framed human demonstrations for manipulation research.

The cited arXiv source contains the full hardware setup, alignment method, policy architecture, experiments, and project link.