Robot world model trained on 15 hours of video generalizes to unseen bodies
jon_barron · x · 2026-07-24
Researchers trained a video world model using only 15 hours of video from a single-arm robot.
- It generalizes zero-shot to unseen embodiments, including orangutans.
- The model also learned an emergent capability not explicitly trained for: given a desired object motion, it can synthesize the robot motion that would produce it.
- The key idea is to represent actions as masked video.
The post highlights a compact but surprisingly general robot-world-model setup, with an emphasis on embodiment transfer and action generation.
Related event: Zero-Shot Robot World Model Trained with 15 Hours of Video(3 posts)→
More from Embodied
- Full teardown: inside Zmotor's 32mm linear actuator for humanoids — MarwaEldiwiny · 2026-10-27
- Full Teardown of a Zmotor Linear Actuator Used in China's Humanoid Robots — MarwaEldiwiny · 2026-10-27
- Video AI agent runs a physics engine, then paints your real room over the blocks — jonstephens85 · 2026-10-27
- World Model Assigns Weight and Friction to Objects, Then Simulates with a Physics Engine — jonstephens85 · 2026-10-27
- First unboxing shots of Apple's foldable "iPhone Duo" surface — chrisfirst · 2026-10-27
- UniAction: unified action conditioning lets one world model handle mixed action types — ZeYanjie · 2026-10-27