Dyna Robotics unveils DYNA-2, a robot foundation model trained on 1 million hours of human video

Dyna Robotics unveiled DYNA-2, a world-action model pretrained on 1 million hours of first-person human video, equivalent to about 170 years. The company says it is the first to demonstrate a “human-to-robot scaling law,” in which robot performance improves predictably using only human video, without robot data. For high-precision manufacturing tasks, it reported that scaling up pretraining alone raised the success rate from 20% to 80%–90%.