ActiveMimic: Egocentric Video Pretraining with Active Perception

ActiveMimic, from Fudan University, reframes egocentric video pretraining for robot manipulation around active perception: instead of treating head-mounted camera motion as noise, it recovers synchronized camera and bimanual wrist trajectories from a single RGB egocentric stream, aligns them in a common reference frame, and encodes them as a unified 27D action space. The method uses off-the-shelf pose models, including SAM-3D-Body, VGGT, and UniDepth, and trains a 3B visual-language prefix plus a 0.6B action expert with flow matching on filtered Ego4D manipulation clips. On a humanoid AGI-BOT G1 across four real-world tasks, ActiveMimic reaches 90.1% on Restocking, 88.9% on Reaching, 91.7% on Finding, and 93.3% on Pouring, outperforming human-video baselines such as MotoVLA and matching or beating π0, a robot-data-pretrained model. Its pretraining corpus contains 2,561 Ego4D episodes, roughly 10 hours at 10 fps. On HOT3D, the recovered labels achieve 78.82% head recovery under the strict tolerance tier, with 65.93% and 61.72% for the left and right wrist. Analysis shows the active-perception capability comes from egocentric pretraining, not robot fine-tuning: on Restocking placement, ActiveMimic scores 24 out of 27, but only 1 out of 27 when the head camera is removed, and camera-motion supervision yields consistently higher head-view/full-view activation overlap in early-to-mid layers.