Folding a restaurant napkin seems simple. At Din Tai Fung, the restaurant chain known for its dumplings, the folded napkin's seam has to sit only a couple of centimeters below the pocket where diners slide their chopsticks. Jason Ma, co-founder and chief scientist of Dyna Robotics, said his company's robots fold thousands of napkins a day for many of the chain's US locations. In his telling, the important part is how many locations use them.
Ma made the case in a conversation on the Latent Space podcast, published October 13, 2026. He called Din Tai Fung Dyna's "biggest scaled customer."
Pilot versus rollout
Many robotics companies talk about deployments and pilots, Ma said. But "what's really hard is not putting your robot in one location or doing a pilot." The difficult step, he argued, comes when a customer decides the robot creates enough value to install it across many of its sites. He said this is what happened with Din Tai Fung, and that to the best of his knowledge it "hasn't really happened" for other companies building similar general-purpose robots.
Din Tai Fung is not Dyna's only customer. Ma said Dyna also folds laundry, mainly towels, for the hotel chain Best Western, work he said every hotel, gym and fitness center needs. Asked which Best Western visitors could go to if they wanted to see the robots, he suggested going to a Din Tai Fung instead.
Why a general model for one chore
The conversation raised an obvious objection: a machine built only for napkins might fold them perfectly, to the millimeter. Why use a general-purpose robot?
Ma's answer was about what napkins are made of. Cloth is a deformable object. It changes shape whenever it is touched, so a napkin or towel can land in infinitely many possible configurations. He said that is hard for a specialized robot to handle. Dyna's approach is to train one broad model and then adapt it to a specific job, and he said this reached production quality faster. The discussion also touched on speed: a custom folding machine has to be designed and built for that one job, while an adaptable robot can be retrained.
Ma said the robots now work at "human parity" in speed, and they do not get tired. He also pointed to consistency. Din Tai Fung's New York and Seattle locations fold napkins slightly differently, he said, and the company wants every restaurant to do it the same way. A robot's model is "just a set of weights," Ma said: the trained numbers that define its behavior. The same weights can be loaded into robots at every site, and when the model improves, Dyna can "update all of them at once."
What the company's reports add
Dyna's own published accounts give a more detailed picture of this work. The figures below come from the company, not from independent measurement.
In an August 2026 deployment account, Dyna said that over an 18-hour restaurant shift its earlier system folded about 35 napkins an hour, with 75% judged acceptable, for roughly 480 usable napkins. The newer Dyna-2 model reportedly folds 95 an hour with 93% acceptance, about 1,590 usable napkins against a requirement of 1,500. That is a little more than three times the earlier output. The same account describes how Dyna monitors its fleet. The robots' work is automatically labeled by step, duration and failure. When one robot slowed down, those labels showed a rise in missed grasps, and the cause turned out to be a worn gripper rather than a problem with the model.
A widely cited reliability figure comes from Dyna's earlier Dyna-1 report, published in June 2025. In that 24-hour napkin run, the robot made more than 850 folds with a 99.4% task success rate and no human intervention in folding. Dyna graded quality separately. Of those folds, 98% reached at least grade three, but 75% met the stricter grade-four-or-five production threshold. The report also said that at customer sites the robot had not seen before, the starting model could fold napkins, but throughput and quality dropped until it received more training on site.
The work around the robot also matters. In a May 2026 essay published by Dyna, York Yang wrote that the company's longest-running deployment had been in daily use for about ten months. He also wrote that reliable operation for customers took weeks to months of on-site engineering, and that this did not initially get faster as Dyna added sites. Yang counted remote diagnostics, maintenance and reusable integrations as part of the product itself.
From napkins to knives
Ma said Din Tai Fung has now signed "a much bigger contract" with Dyna. The chain also wants the robots to cut food, he said, given how many vegetables each restaurant gets through every day. Asked whether that was the second use case, he said it was one of the uses the chain is considering. When the conversation turned to robots eventually making the dumplings, Ma called that "the final boss."
Ma said Dyna is not aiming to master any single task. Each job, he said, serves as validation of a general platform and as "a wedge" for learning how to build everything better. Dyna has already deployed tens of robots in laundry and restaurants, he said. Its goal for next year is to deploy hundreds.