Open-HumanScale pushes egocentric video pretraining to 50,000 hours, and robot success climbs log-linearly with the hours

Egocentric human video now serves as a scalable pretraining source for robot policies, with scale pushed from 100 hours to 50,000 hours and downstream success rising steadily with pretraining hours. Earlier work in this line measured human-data scaling mainly through validation action loss or zero-shot action prediction, without a systematic closed-loop comparison on task success rates. Under isolated attention, RoboTwin randomized-scene success rises from 8.4% at 100 hours to 28.7% at 10,000 hours, against 2.1% without egocentric pretraining; scene-randomized success rises from 25.2% to 58.1%, against 16.2% without pretraining.

The recipe maps human manipulation trajectories from different sources into one bimanual end-effector action space, so heterogeneous data can jointly train a single world-action model. Each egocentric dataset estimates actions through a different pipeline of hand keypoints and wrist poses, so raw values lack a consistent physical meaning and direct mixing weakens supervision. On a 1,000-hour corpus built from EgoSuite-Open100K, HumanNet, EgoDex and EgoVerse, the aligned model reaches 26.4% randomized-scene success versus 18.0% unaligned, and the aligned model also shows lower action loss during pretraining.

Jointly pretraining on human and real-robot data with visual-goal conditioning lets the model execute tasks from a goal image without parameter updates. Models pretrained on human data alone carry an embodiment gap and still need task-specific demonstration finetuning before deployment. Adding 45,900 hours of human data to a fixed 14,600-hour robot corpus raises mean zero-shot success across 20 real-robot tasks from 20.0% to 30.5%; after post-training, HumanWAM-Goal reaches 92.70% on RoboTwin 2.0 Clean2Clean and 84.60% on Clean2Random.

The demonstrated scope centers on bimanual tabletop manipulation: simulation evaluation uses RoboTwin 2.0 and RoboDojo; HumanWAM-Vanilla's real-robot experiments on Cobot Magic cover paper cup stacking, towel folding, drawer storage and desk tidying, while HumanWAM-Goal is evaluated on 20 zero-shot tasks on the Astribot S1 wheeled mobile manipulator. Readers should treat it as scale evidence for an egocentric pretraining recipe, not a general conclusion across all robot embodiments.

The log-linear relationship is an empirical fit, with goodness-of-fit reported separately for the two attention configurations, not a theoretical guarantee. The purely human-pretrained model still trails some compared models on RoboTwin randomized scenes, which the authors partly attribute to a basic world-action architecture without specialized modules. On real robots, hard-task success plateaus under standard data-proportional sampling and improves only when the human-to-robot sampling ratio is rebalanced.