Robot data is expensive and robot hours do not scale like text tokens. The workaround most labs reach for is human video: egocentric footage of hands doing things, plentiful and free, but raw in the sense that a few hundred hours of uncurated YouTube does not transfer into a policy that grips anything reliably.
Open-HumanScale, posted to arXiv on 4 November, is an attempt to make that transfer systematic — and to release the whole thing.
The Move
The work anchors on Open-HumanScale-50K, a curated egocentric video corpus of 50,000 hours, densely annotated with hand motion and natural language. The authors use it to investigate pure human pretraining through closed-loop simulation evaluation, and they report an empirical log-linear scaling law linking pretraining data hours to downstream robot success rate — the kind of curve that lets you predict what 100,000 hours buys before you collect it.
Two models come out of it:
- HumanWAM-Vanilla, a world-action model pretrained only on the 50k hours of human video. After robot post-training it hits 92.08% on RoboTwin 2.0 Clean2Clean and 68.88% on Clean2Random, with 14.17% mean success and a 19.00 mean progress score on the harder RoboDojo
- HumanWAM-Goal, a visual-goal-conditioned world-action model trained on hybrid human+robot data, pushing RoboTwin 2.0 to 92.70% Clean2Clean and jumping Clean2Random to 84.60%
On physical robots HumanWAM-Goal shows emergent embodied in-context learning — adapting its manipulation to unseen visual goals with no parameter updates.
What It Buys
The gap between 68.88% and 84.60% on Clean2Random is the interesting part: that benchmark varies the objects, so it measures generalization rather than memorization, and it is where most policies fall apart.
The release matters as much as the numbers. The team open-sources the corpus, the curation and pretraining pipelines, the pretrained and post-trained checkpoints, inference code and the simulation evaluation suites — the full stack, not just the weights.
Why It Matters
If the log-linear scaling law holds outside this lab’s setup, “how much human video do we need” becomes a budget question rather than an open research program. And with the corpus and recipe public, the 50,000-hour bar is a starting line others can actually cross.
Paper: arXiv:2611.06817