Open-HumanScale Scales Embodied Learning with 50,000 Hours of Human Video

Robots can learn manipulation skills from demonstrations, but collecting large amounts of high-quality robot data remains expensive. Human videos offer a much larger source of interaction experience, yet they do not directly specify the actions a particular robot should execute. Open-HumanScale addresses this gap with an open data and training recipe for using egocentric human experience in embodied foundation models.

The project introduces the Open-HumanScale-50K corpus, a curated collection containing 50,000 hours of egocentric video. The data is densely annotated with hand motion and natural language, giving a model access not only to visual observations but also to signals about manipulation and task descriptions. The authors frame curation and training design as central parts of scaling: increasing the raw number of hours alone does not guarantee more general robot policies.

Key points

• A measured human-video scaling study. Using closed-loop simulation evaluations, the authors examine how pretraining hours relate to downstream robot performance. Their results indicate an empirical log-linear relationship, with broader data diversity strengthening the benefits of scaling. Human pretraining also improves the efficiency of learning from robot demonstrations.
• HumanWAM-Vanilla isolates human experience. This world-action model is pretrained exclusively on 50,000 hours of egocentric human video and later post-trained with robot data. It reaches 92.08% success on RoboTwin 2.0 Clean2Clean and 68.88% on Clean2Random. On RoboDojo, it records a 14.17% mean success rate and a 19.00 mean progress score.
• HumanWAM-Goal combines two forms of experience. This model jointly uses human and robot data and introduces visual goal conditioning. After post-training, it achieves 92.70% on RoboTwin 2.0 Clean2Clean and 84.60% on Clean2Random, with an average of 88.65% across those settings.
• Goal-conditioned adaptation. The released findings describe zero-shot transfer across simulation and physical benchmarks. On physical robots, the model can adapt manipulation behavior to unseen visual goals through in-context conditioning, without updating parameters for each target task.

The important contribution is therefore broader than a new checkpoint or a larger video collection. Open-HumanScale makes the scaling question more concrete: which parts of human experience transfer to robots, how much robot data is still required, and how can visual goals guide behavior without task-specific optimization? Egocentric footage may not provide an exact executable trajectory for every robot embodiment, but it can encode useful priors about hand-object contact, motion patterns, and task structure.

The project also supports reproducibility by releasing the corpus, curation and pretraining pipelines, HumanWAM checkpoints, inference code, and simulation evaluation suites. Documentation and uploads are still being updated, so users should inspect the released configurations before comparing results. Even with that caveat, the work offers a practical public framework for studying whether human experience can scale alongside robot data in embodied AI.

Source: Hugging Face Daily Papers