30,000 Hours of Egocentric Video Won't Teach World Models Object Interaction
A study by Toyota Research and Jiahua Dong's team trained a world model on 30,012 hours of egocentric video. The model learned hand movements but not the physical consequences on touched objects, and scaling data from 300 to 30,000 hours did not close the gap.
2026-10-13 ~ 2026-10-14 · 2 related posts
- 30,000 hours of egocentric video: world model learns hands, not object outcomes — xuanalogue · 2026-10-13
- 30,000 Hours of Egocentric Video Teach a World Model Hands, Not Object Physics — mangahomanga · 2026-10-14