30,000 hours of egocentric video: world model learns hands, not object outcomes
xuanalogue · x · 2026-10-13
Training a world model on 30,000 hours of first-person video, Jiahua Dong's team found it learned where hands go but far less about what happens to objects they touch. Scaling from 300 to 30,000 hours didn't narrow the gap, and a 4× larger model barely helped. Tonghe Zhang argues this quantitatively disproves claims that scaling video data solves robotics: POMDPs contain states that can never be decoded from observations.
More from Embodied
- Spranic Demos Spatial Memory That Turns Building Scans Into Queryable Asset Records — AI Engineer · 2026-10-13
- Grounded Action Model Bets on 3D Grounding as the Foundation for Robotics, Fully Open-Sourced — joemeno · 2026-10-13
- Podcast: why humanoids get to defense before your home does — audrow · 2026-10-13
- Humanoid robots at scale 20-30 years away: copper, lithium shortages and boomer retirements reshape the AI jobs debate — JHochderffer · 2026-10-13
- Shield AI's Hivemind cut Honeywell's autonomy development from a year to six weeks — BrettKrieger12 · 2026-10-13
- Perplexity design chief jokes about wearing Apple HomeView like Flava Flav — davidhoang · 2026-10-13