Dyna Robotics deploys Dyna-2.1 robots folding napkins at Din Tai Fung with 99.4% success rate
Dyna Robotics co-founder and chief scientist Jason Ma detailed the company's Dyna-2 world-action model and its physical deployment Dyna-2.1 on the Latent Space podcast. The model was pre-trained on over one million hours of human egocentric video, using a diffusion transformer architecture with flow matching for real-time control. Ma said the system achieves a 99.4% success rate folding…
Key points
- Dyna-2 pre-trained on over one million hours of human video using diffusion transformer with flow matching
- Dyna-2.1 robots deployed at Din Tai Fung and Best Western, achieving 99.4% napkin-folding success over 24 hours
- New tasks learned from 45 minutes of demonstrations; Dyna builds custom hardware, actuators, and on-device inference stack
Model page: GPT-6 →
Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Latent Space · 13 October 2026
Loading the full article…
This text was published by Latent Space and written by Latent.Space. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Robotics & Physical AI
All →- Omron buys OnLogic for $675M to expand physical AI hardware · 1 src
- Brazilian Developer releases PS5 mod to run Xbox Cloud Gaming · 1 src
- Hello Robot and A&K Robotics to discuss indoor robot navigation at RoboBusiness 2026 · 1 src
- UCLA builds PhoneBot humanoid from smartphone and $400 hardware · 1 src
- Atomic Machines raises $250M to build tiny machines from code with AI · 1 src
Comments
via GitHub Discussions