Dyna ships Dyna-2.1, claims hour-long robot autonomy

Dyna Robotics followed its Dyna-2 world-action model, announced last month, with Dyna-2.1, a physical robot agent it says achieves reliable super long-horizon whole-body autonomy. The company pairs that agentic system with new semi-humanoid hardware rather than off-the-shelf humanoids; co-founder and chief scientist Jason Ma has explained that choice as a hardware decision, and the episode touches on why Dyna builds its own hardware at all.

The underlying model is the other half of the story. Dyna-2 is a diffusion-transformer world-action model trained on more than one million hours of human video — the company says it found new scaling laws on human data across four orders of magnitude. In the episode, Ma describes the robotics data problem plainly: unlike language models, robotics has no internet, so a single company cannot collect enough data on its own. The recipe he outlines combines large egocentric human video to give the model a broad prior with a smaller amount of robot-collected data for low-level dexterity, which is why Dyna invited data vendors to reach out through an email address listed in its Dyna-2 blog post.