Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics

Dyna Robotics founder Jason Ma describes Dyna-2 as a robot model that predicts what a scene will look like next and uses that prediction to guide physical movements. Its inputs include recent camera images, joint positions and a language instruction; its outputs include robot actions and predicted future video or a representation of it. This connects an instruction to an anticipated physical outcome. Company interview.

Adapting the model to a task still involves demonstrations on the robot. Ma describes collecting tens of minutes to a few hours of human-operated examples, testing the trained model in the same setup, and collecting more data where it struggles. That is a concrete adaptation process, with real-world trials determining what needs improvement. Training discussion.