Peking University-led team open-sources 50K-hour recipe for embodied foundation models
DAGroup-PKU · hf · 2026-11-06
A PKU/NUS/MIT/UCSB/NVIDIA team releases Open-HumanScale, the first fully open recipe for scaling egocentric video pretraining for embodied foundation models:
- Open-HumanScale-50K: a curated 50,000-hour egocentric video corpus densely annotated with hand motion and natural language, open-sourced along with curation/pretraining pipelines, checkpoints, and simulation eval suites.
- Closed-loop simulation evaluations reveal an empirical log-linear scaling law between pretraining hours and downstream robot success rates.
- HumanWAM-Vanilla, pretrained purely on human video, reaches 92.08% on RoboTwin 2.0 Clean2Clean and 68.88% on Clean2Random after robot post-training.
- HumanWAM-Goal, with joint human-robot pretraining and visual-goal conditioning, achieves broad zero-shot generalization and a new SOTA of 84.60% on Clean2Random.
- On physical robots it shows emergent embodied in-context learning, adapting to unseen visual goals without parameter updates.
More from Embodied
- Xinhua hires two UBTECH UWORLD U1 humanoid robots as news anchors — CyberRobooo · 2026-11-06
- SLA scales spatial-language-action models via continual pre-training with frozen 3D alignment — HKUSTGZ · 2026-11-06
- KAIST's DreamFree distills video diffusion representations for real-time robot control — kaist-ai · 2026-11-06
- Wayve details 'Learned Safety' approach to validate end-to-end AVs at scale — Jamie_Shotton · 2026-11-06
- Why hasn't anyone hooked an AI agent to a swarm of remote-controlled cyborg cockroaches? — Sauers_ · 2026-11-06
- Physical AI meetup in India: 4 robotics founders on real factory deployments and eval loops — carrycooldude · 2026-11-06