Dyna Robotics says its Dyna-2.1 physical agent can execute hour-long, multi-step jobs and recover from errors rather than stopping at the first failed action. In one company demonstration, a robot folded restaurant napkins for 24 hours at a 99.4 percent success rate, including recovering after it knocked over the stack.
The agent is built around Dyna-2, a “world-action” model trained on more than one million hours of human video. It consumes camera images, robot joint positions and language instructions, then predicts both future visual states and physical actions. Dyna argues that this future prediction helps the system plan smooth movements and generalize to unfamiliar settings.
Chief scientist Jason Ma told Latent Space that new tasks can require as little as 45 minutes of demonstrations. The company is already deploying robots to fold thousands of napkins at restaurant chain Din Tai Fung, moving the technology beyond a laboratory demonstration.
The claims are based on Dyna’s own tests and an interview rather than an independent evaluation. Ma also identified persistent limits: robotics lacks an internet-scale action dataset, dexterous hands can overheat, and real-time control requires model distillation and on-device inference. Dyna chose wheeled semi-humanoid hardware instead of a bipedal design, prioritizing reliable commercial work over a humanlike form.