Dyna-2: Robot Task Improvement via 1M Hours of Human Video Pre-training
alifcoder · x · 2026-08-11
Dyna Robotics introduces Dyna-2, a world-action model pre-trained on over 1 million hours of human first-person video. The research reveals that this pre-training significantly improves robot task performance even without any robot trajectory data.
The study also uncovers several new scaling laws: world-action models exhibit scaling laws on human data across four orders of magnitude; this implies a scaling law on unseen robot data; and both data and objectives are crucial, with world modeling on video data being essential for cross-embodiment transfer.
More from Embodied
- Kyle Vogt's Bot Company Sued for Turning Airbnb into Robot Lab — MarwaEldiwiny · 2026-08-11
- Dyna-2 Model Transfers to WUJI2 Dex Hand, Adapts in 13 Minutes of Data — JasonMa2020 · 2026-08-11
- MIT's GeoPT Model Cuts Simulation Data Needs by 60% with Built-in Physics — nordicinst · 2026-08-11
- 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