Dyna-2: Robot Foundation Model Trained on 1M+ Hours of Human Data Reveals New Scaling Law
JasonMa2020 · x · 2026-08-11
Introduces Dyna-2, the first robot foundation model trained on over 1 million hours of human data.
- Cross-embodiment transfer scaling law: Training on increasing amounts of human data not only improves prediction on held-out human data but also boosts performance on unseen robot data.
- Real-world validation: This transfer scaling law successfully translates to on-robot performance across 3 different robot platforms.
- Key insights: Both training objectives and data matter greatly for this emergence, positioning video as a new scaling axis for physical AI.
- Capabilities: The post dissects its skills, including language following via world modeling, enhanced robustness/precision, zero-shot customer site deployment, and video generation.
The author notes this represents a fundamentally different and highly promising path for robot foundation models.
Related event: Dyna-2: Million-Hour Human Video Pretraining Unlocks Embodied Scaling Laws(13 posts)→
More from Embodied
- MIT and Tsinghua introduce GeoPT: physics as AI's third modality — MIT News AI · 2026-08-11
- Amazon Gutting Nova AI Models, Pivoting Hard to Robotics — max_paperclips · 2026-08-11
- Big Tech Robotics Engineers Make $400K+, Startups Face a Math Problem — ATTlKA · 2026-08-11
- Next-Gen BCI: Could Nanoparticles Sprayed Up Your Nose Replace Brain Surgery? — Chris_Armstrong · 2026-08-11
- Meta Ray-Bans Roasted with 'Not My Problem' Filter Meme — jonippolito · 2026-08-11
- MemoMind Smart Glasses Exceed Kickstarter Goal 100x, Raise ~$9M — chrisgrayson · 2026-08-11