Open-HumanScale: Scaling Human and Robot Experience for Embodied AI

arxiv.org1 min read

Researchers introduce Open-HumanScale, an open recipe for scaling egocentric pretraining to 50,000 hours to improve embodied foundation models. By leveraging the Open-HumanScale-50K corpus and developing world-action models such as HumanWAM-Vanilla and HumanWAM-Goal, the approach achieves state-of-the-art success rates on robot manipulation benchmarks and enables emergent embodied in-context learning. Open article →