Open-HumanScale Scales Human Video for Robot Learning
Original title: Open-HumanScale: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models
Open-HumanScale-50K corpus and HumanWAM models pretrained on 50k hours of egocentric video achieve high success rates on robot benchmarks.
Egocentric human video offers scalable manipulation priors for embodied AI. Open-HumanScale provides a fully open recipe and curated 50K-hour corpus annotated with hand motion and language. Pure human pretraining follows a log-linear scaling law to downstream success. Hybrid human-robot pretraining with HumanWAM-Goal enables zero-shot generalization on physical robots.
- 50k hours of egocentric video pretraining
- HumanWAM-Vanilla: 92.08% on RoboTwin Clean2Clean
- HumanWAM-Goal: 84.60% on RoboTwin Clean2Random
- Emergent in-context learning on physical robots
- Corpus and checkpoints open sourced
From the original article
1 min read at huggingface.co
Egocentric human video offers a scalable source of manipulation priors for robot learning, yet translating raw data scale into generalizable robot policies demands rigorous data curation and training recipes.
By Juncheng Ma, Jiaxin Li, Jiankai Tu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.