Glonce

Open-HumanScale releases 50,000-hour egocentric video recipe for robot models

researchThe Glonce editorsHugging Face🔥63Glonce rating63Impact50Novelty65Relevance100Surprise35Players15Credibility70
Open-HumanScale releases 50,000-hour egocentric video recipe for robot models

Open-HumanScale is a preprint on teaching robots from egocentric video, meaning footage filmed from the human's own point of view. The authors start from the idea that such video is a scalable source of manipulation priors for robot learning, but argue that turning raw data scale into policies that generalise takes careful data curation and training recipes. They present Open-HumanScale as the first fully open recipe for scaling egocentric pretraining to 50,000 hours.

The anchor of the recipe is Open-HumanScale-50K, a curated egocentric video corpus densely annotated with hand motion and natural language. Using it, the authors study pure human pretraining through closed-loop simulation evaluations and report an empirical log-linear scaling law that links the hours of pretraining data to downstream robot success rates.

Building on that finding, they train HumanWAM-Vanilla, a world-action model pretrained exclusively on 50k hours of egocentric human video. After post-training on robot data, it reaches 92.08% success on RoboTwin 2.0 Clean2Clean and 68.88% on Clean2Random. On RoboDojo it gets a 14.17% mean success rate and a 19.00 mean progress score.

To push further, the authors test joint human-robot pretraining with HumanWAM-Goal, a visual-goal-conditioned world-action model. They say this hybrid data scaling unlocks broad zero-shot generalisation across simulation and physical benchmarks and sets new state-of-the-art post-trained RoboTwin 2.0 results: 92.70% on Clean2Clean and 84.60% on Clean2Random, which they call a substantial leap. On physical robots, HumanWAM-Goal is reported to show emergent embodied in-context learning: it changes its manipulation behaviour to match unseen visual goals without any parameter updates.

For reproducibility, the authors open-source the Open-HumanScale-50K corpus, the data curation and pretraining pipelines, the pretrained and post-trained HumanWAM checkpoints, the inference code and the simulation evaluation suites.

Key facts

  • Open-HumanScale is billed by its authors as the first fully open recipe for scaling egocentric pretraining to 50,000 hours, built around the Open-HumanScale-50K corpus annotated with hand motion and natural language.
  • In simulation tests of pure human pretraining, the authors report an empirical log-linear scaling law between pretraining data hours and downstream robot success rates.
  • HumanWAM-Vanilla, pretrained only on human video, reaches 92.08% on RoboTwin 2.0 Clean2Clean and 68.88% on Clean2Random after robot post-training.
  • HumanWAM-Goal, pretrained on human and robot data jointly, reports 92.70% on Clean2Clean and 84.60% on Clean2Random, which the authors call new state-of-the-art results on RoboTwin 2.0.
  • The corpus, curation and pretraining pipelines, checkpoints, inference code and simulation evaluation suites are all open-sourced.

Why it matters

Robot learning is short of large, varied manipulation data, and human first-person video is an attractive substitute because it can be gathered at scale. This paper puts a number on that bet: 50,000 hours of annotated egocentric video, plus a reported log-linear relationship between data hours and robot success. If that trend holds up, it gives teams a way to estimate what more human video would buy. The authors also call the recipe fully open, which is rare for work at this scale.

Who it affects

Embodied AI and robotics researchers who train manipulation policies, and groups that want to reproduce or extend large-scale pretraining without collecting their own corpus. Teams that benchmark on RoboTwin 2.0 or RoboDojo now have new reported numbers to compare against.

How to use it

The authors say they release the Open-HumanScale-50K corpus, the data curation and pretraining pipelines, pretrained and post-trained HumanWAM checkpoints, inference code and simulation evaluation suites. In practice that means a researcher could reuse the data, rerun the pipelines, start from the checkpoints, or run the simulation suites to compare against the reported scores. No release date, license, or download link is given for the corpus or checkpoints.

How solid is it

This is a preprint, and the results are the authors' own reports. Numbers are specific: 92.08% and 68.88% on RoboTwin 2.0 Clean2Clean and Clean2Random for HumanWAM-Vanilla, and 92.70% and 84.60% for HumanWAM-Goal. The state-of-the-art claim rests on the authors' comparison, and the abstract gives no baseline numbers or prior state-of-the-art values, so the size of the improvement over earlier methods is not stated. The slope or coefficients of the scaling law are not given either. The physical-robot findings, including the in-context learning behaviour, are described only in general terms.

Risks and caveats

The abstract does not say which physical robots or how many real-world tasks were used, nor the model sizes or compute behind pretraining. The 14.17% RoboDojo mean success rate and 19.00 mean progress score are given for HumanWAM-Vanilla only, not HumanWAM-Goal. The source of the egocentric videos is not stated. The abstract also names no authors or institutions. The scaling law comes from simulation evaluations of pure human pretraining, so it should not be read as a guarantee for real-world robots.

“the first fully open recipe for scaling egocentric pretraining to 50,000 hours”

— Open-HumanScale abstract