30,000 Hours of Egocentric Video Teach a World Model Hands, Not Object Physics
mangahomanga · x · 2026-10-14
- New Toyota Research-affiliated work trains egocentric world models (Cosmos 3-based, 4B/16B/64B) on 30,012 hours of first-person manipulation video: 1.15M clips, 32,051 unique tasks, 3.24B frames.
- Key finding: agent fidelity (where hands go) saturates quickly, but object fidelity (what happens to things hands touch) lags far behind. Scaling data 300h→30k hours didn't close the gap, and a 4× larger model barely helped.
- Conclusion: human video lacks force annotations and satisfying actions, an inherent data limitation scaling can't fix. Paper page live; code and benchmark coming soon.
More from Embodied
- China reportedly shipped 97% of the world's humanoid robots in H1 2026 — Polymarket · 2026-10-14
- Physical Intelligence opens Pi API beta, one model adapting robots from strawberry harvesting to server assembly — svlevine · 2026-10-14
- China's Tiangong Ultra robot beat Bolt's 100m record — and exposes America's robotics supply-chain dilemma — mattsheehan88 · 2026-10-14
- GoDrift shows off realtime sim2real robot demo; 2D plane eases reaction demands — jonstephens85 · 2026-10-14
- Microsoft science president Peter Lee leaves to become Mercedes F1 team's AI advisor — lazowska · 2026-10-14
- NVIDIA releases GR00T N1.7 AppleToPlate checkpoint-15000 on Hugging Face — nvidia · 2026-10-14