Open-HumanScale: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models

Open-HumanScale. An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models. By Juncheng Ma and thirty others, from Peking University, the National University of Singapore, and three other institutions. Published on the fourth of November.

Video of people using their hands, filmed from their own point of view, is a scalable source of manipulation priors for robot learning. There is a great deal of it. Having a lot of raw data, though, is not the same as having robot policies that generalize. Turning scale into something a robot can use demands rigorous data curation and a training recipe, and those are the parts that tend to stay hidden.

Open-HumanScale is presented as the first fully open recipe for scaling egocentric pretraining to fifty thousand hours. At its centre is a corpus called Open-HumanScale-50K: curated egocentric video, densely annotated with hand motion and with natural language. The recipe is the corpus together with the curation and pretraining pipelines that produce models from it.

With the corpus in hand, the authors first study pure human pretraining, measured through closed-loop simulation evaluations. That study uncovers an empirical log-linear scaling law: as the hours of pretraining data grow, downstream robot success rates rise along a log-linear curve. Those findings then ground a model. HumanWAM-Vanilla is a world-action model pretrained exclusively on fifty thousand hours of egocentric human video, with no robot data in pretraining.

A second model pushes further. HumanWAM-Goal is a world-action model conditioned on visual goals, and it is pretrained jointly on human and robot data. That hybrid data scaling is what the authors credit with broad zero-shot generalization, across both simulation and physical benchmarks.

After robot post-training, HumanWAM-Vanilla reaches ninety-two point zero eight percent success on RoboTwin 2.0 in the Clean2Clean setting, and sixty-eight point eight eight percent in Clean2Random. On RoboDojo it records a fourteen point one seven percent mean success rate and a mean progress score of nineteen.

HumanWAM-Goal moves the post-trained RoboTwin 2.0 results to new state of the art levels: ninety-two point seven percent on Clean2Clean, and a substantial leap on Clean2Random, to eighty-four point six percent. On physical robots it also shows what the authors call emergent embodied in-context learning. Faced with visual goals it has not seen, it adapts its manipulation behaviour on the fly, without any parameter updates.

Openness is the point the work keeps returning to. To support reproducible scaling research in embodied AI, the release covers the Open-HumanScale-50K corpus, the data curation and pretraining pipelines, the pretrained and post-trained HumanWAM checkpoints, and the simulation evaluation suites. That puts the whole chain in the open, from raw human video to a robot's success rate, so that the scaling law itself can be checked by others rather than taken on trust.