Dyna-2.1 Physical Agent Achieves Hour-Scale Autonomous Workflows

Dyna-2 is a world-action model trained on over one million hours of human video, built on an encoder-only diffusion Transformer with flow-matching to speed up sampling. Its core idea is to first predict future video states and then generate robot actions from them, akin to chain-of-thought reasoning in visual and physical space, improving dexterity, robustness, and generalization to unseen environments. Inputs include multi-second video history, joint encoder angles, and language instructions; outputs are executable actions and future video representations. Dyna-2.1 is described as the first physical agent to achieve reliable ultra-long-horizon whole-body autonomy. In official demos, the robot folded restaurant napkins continuously for 24 hours with a 99.4% success rate and recovered on its own after knocking over a stack. Dyna chose custom wheeled semi-humanoid hardware over off-the-shelf humanoids, arguing that on-device inference, model distillation, and custom hardware are prerequisites for reliable deployment; dexterous hand overheating remains a hardware bottleneck. The team says robots can learn new tasks from as little as 45 minutes of demonstration and has deployed at scale with real customers such as Din Tai Fung. For teams shipping robots, this approach shows that data scale and self-generated training data (synthetic environments, simulation) matter more than hand-designed specialized solutions; the route difference between world-action models and vision-language-action models also directly shapes model selection and data collection strategy.