Dyna Robotics Unveils DYNA-2 Robot Foundation Model

Dyna Robotics announced in a press release DYNA-2, a robot foundation model trained on more than 1 million hours of egocentric human video. The company said the model uses video data rather than manually collected robot action data to train robots for physical tasks.

DYNA-2 uses a dual next frame and next action world modeling architecture. Dyna Robotics said the model can transfer learning across stationary robot arms, humanoid prototypes, and dexterous robotic hands.

In manufacturing tasks cited by the company, DYNA-2 raised task success rates from 20% to 80% to 90% through larger pre training scale, without changes to post training data. Across 15 benchmark tasks, the company said policies trained with more human video data performed better.

Dyna Robotics also compared DYNA-2 with DYNA-1 in physical evaluations using matched training steps and datasets. The company said DYNA-2 completed tasks 1.55 times more often, reached an 87% pass rate at one customer deployment compared with 46% for DYNA-1, and used 13 minutes of local data to command a pair of five fingered robot hands to open a bottle cap.