Dyna-2: Human Data Scaling Law Predictably Improves Zero-Shot Robot Performance

chris_j_paxton · x · 2026-08-11

DynaRobotics shared a key finding from the Dyna-2 research: the human data scaling law implies a scaling law on never-seen robot data.

The research shows that scaling human data for pre-training predictably improves zero-shot offline evaluation on robot data. Their blog also demonstrates that this scaling successfully transfers to on-robot evaluation through post-training on diverse robot embodiments.

Related event: Dyna-2 Trained on 1M Hours of Human Video Reveals New Embodied Scaling Laws(11 posts)→

Original post →

More from Embodied

Embodied channel →