On the Latent Space podcast, Dyna Robotics co-founder and chief scientist Jason Ma detailed Dyna-2, a "world-action model" trained on more than one million hours of human video. The model is designed to predict future video frames while also outputting robot actions, an approach the company distinguishes from vision-language-action models. Ma argues that predicting the physical future is key to general-purpose robotics, and that data scaling, not hand-engineered solutions, will drive progress.
Following Dyna-2, the company released Dyna-2.1, a physical agent built on custom wheeled semi-humanoid hardware. Dyna says the system combines the model with an agentic harness to handle long, multi-step workflows. In one evaluation, the robot folded restaurant napkins for 24 hours straight with a 99.4% success rate, including recovering after knocking over the stack. The company also says it has deployed robots at Din Tai Fung that fold thousands of napkins daily.
Ma also discussed remaining bottlenecks: dexterous hands still overheat, sensory data is incomplete, and robotics lacks an internet-scale dataset. Dyna builds its own hardware rather than using off-the-shelf humanoids, and it says new tasks can be learned from as little as 45 minutes of demonstrations. The episode frames these results as early evidence for a "bitter lesson" approach to robotics, scaling data instead of specialized engineering.