Dyna Robotics

Dyna Robotics builds robots and models for physical tasks such as handling laundry and folding restaurant napkins. Its chief scientist Jason Ma says Dyna-2 was trained on more than one million hours of human data, then adapted with robot demonstrations; he reports that the company’s robots fold napkins at multiple Din Tai Fung locations.[1]

Dyna calls its system a world-action model. Ma describes inputs including recent camera images, robot joint positions and a language instruction. The model can predict future video or a representation of it as well as output actions for the robot. In his account, the predicted future helps guide physical movement; this is an architectural approach, not a guarantee of success in every environment.[1]

Human video supplies varied examples of how objects and people move, but it does not directly specify the actions of a robot with a different body. Dyna therefore adds robot data and examples of the target task. The hosts described a chopping demonstration using 45 minutes of data; Ma says other tasks require tens of minutes or a couple of hours, followed by testing on the physical setup and collecting more data where needed. These are examples, not a universal training-time promise.[1]

The team’s semi-humanoid design has an upper body with arms and a moving waist on a wheeled base. Ma says the torso enables movements that resemble the human data used in pretraining, while wheels serve near-term practical and stability considerations. Dyna buys mature components such as lidar and cameras while pursuing control over other parts it considers important to performance.[1]

Ma reports that Dyna robots fold towels for Best Western and thousands of napkins per day across multiple Din Tai Fung locations. He describes proposed food cutting as a possible next use case. These are company-reported deployments and plans; the interview does not supply an independent performance audit or establish that every site uses the robots.[1]

Longer tasks remain difficult. Ma describes a laundry workflow in the hour range, but says generalizing such work depends on both data and hardware. Dexterous hands can overheat, and vision-only recordings may miss tactile and audio information needed for some tasks. Dyna’s approach is to test models on real hardware and iterate with further examples, while treating broader autonomy as an ongoing goal.[1]