arXiv 2611.06817
Open-HumanScale: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models
Robotics14 upvotes
Authors: Juncheng Ma, Jiaxin Li, Jiankai Tu, Bo Liang, Guodong Lv, Yufan Deng, Junzhe Tang, Liangzhu Yang +23
Overview
The authors release Open-HumanScale, an open recipe for scaling egocentric human video pretraining to 50,000 hours, built around a curated corpus (Open-HumanScale-50K) annotated with hand motion and language. They train HumanWAM-Vanilla, a world–action model pretrained only on human video, and HumanWAM-Goal, a visual-goal-conditioned variant trained with joint human–robot data.
Introduction
Raw egocentric human video could supply useful manipulation priors for robots, but turning that data scale into generalizable robot policies requires careful data curation and training recipes.
Results
Post-trained HumanWAM-Goal reaches 92.70% on RoboTwin 2.0 Clean2Clean and 84.60% on Clean2Random, new state-of-the-art levels. Pretraining on human video alone shows an empirical log-linear scaling law between pretraining hours and robot success rates.
Conclusion
By open-sourcing the corpus, pipelines, checkpoints, and evaluation suites, this work gives the embodied AI community a reproducible basis for studying scaling of human video pretraining.