New Robot Scaling Law: Pre-training on 1M Hours of Egocentric Video Boosts Embodied AI

Scobleizer · x · 2026-08-12

The tech community is buzzing about a potential new "robot scaling law." Research suggests that robot intelligence can effectively scale with massive amounts of video data.

Using the Dyna-2 model as an example, the model was pre-trained on 1 million hours of egocentric video without incorporating any actual robot data. The results demonstrate a clear scaling behavior: as video data scales from 1K to 1M hours, robot performance improves correspondingly. This implies that robotics may finally have its own scaling law.

Related event: Dyna-2 Model: Million-Hour Human Video Pretraining Unlocks Embodied AI Scaling Laws(42 posts)→

Original post →

More from Embodied

Embodied channel →