Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Dyna Robotics details Dyna-2, a world-action model trained on a million hours and deployed on working robots.
Dyna Robotics co-founder Jason Ma discussed Dyna-2, a world-action model pretrained on more than one million hours of human video that predicts future frames and produces robot actions. The follow-on agent, Dyna-2.1, reportedly reached a 99.4% success rate folding restaurant napkins for 24 hours and runs in production at customers including Din Tai Fung and Best Western. Ma described learning new tasks from roughly 45 minutes of demonstrations using diffusion transformers and flow matching, with on-device inference on custom wheeled hardware.
- Dyna-2 was trained on over one million hours of human video.
- The model predicts future video and outputs robot actions.
- It reportedly folded napkins at 99.4% success for 24 hours.
- New tasks are learned from about 45 minutes of demonstrations.
- Robots are deployed at Din Tai Fung and Best Western.
- Products
- Taku
- Organizations
- Dyna Robotics
From the original article
15 min read at latent.spaceOur first robotics pod with Dyna’s Chief Scientist on their big bet on World Action Models & the Next Frontier of Physical AI
latent.space
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.