Dyna Robotics turns napkin folding into a Din Tai Fung rollout
Dyna says Dyna-2 can produce 1,590 table-ready napkins per shift, enough to move from a 2025 pilot into a restaurant-network rollout.
By Ryan Merket · Published
Primary source: X
Why it matters
Dyna Robotics has attached a foundation-model claim to a customer workload and an economic threshold. Scaling the Din Tai Fung rollout will show whether that performance survives across sites without custom engineering swallowing the savings.
Dyna Robotics founders Lindon Gao, York Yang and Jason Ma are turning a napkin-folding robot into a rollout across Din Tai Fung's restaurant network, after the robotics startup said Thursday that its Dyna-2 system had passed the restaurant group's return-on-investment threshold.
https://x.com/DynaRobotics/status/2093039146357178475
The Redwood City, California company said in a thread on X that a robot running Dyna-2 can produce 1,590 table-ready napkins during an 18-hour shift. Din Tai Fung requires 1,500 from each machine to keep a dining room supplied, according to Dyna Robotics.
That figure is Dyna Robotics' central evidence for its ROI claim. The company's public case rests on output and quality rather than customer financials: Dyna-2 folds 95 napkins an hour, with 93% meeting Din Tai Fung's quality standard. A year earlier, Dyna-1 produced about 480 acceptable napkins per shift, folding roughly 35 an hour with a 75% acceptance rate.
The result appears narrow by design. Napkin folding gives Dyna Robotics a repetitive, stationary task with a clear quality test, a known daily workload and few reasons to build an expensive humanoid body around the underlying model. It also gives the founders something robotics demonstrations routinely avoid: a customer-defined production target that has to be met every day.
Din Tai Fung is a demanding place to test that argument. The restaurant chain averaged $27.4 million in annual sales per US location in 2024, the highest average unit volume among US restaurant chains tracked by Technomic, Restaurant Business reported. High throughput leaves little room for a machine that needs frequent staff intervention.
The last few inches of automation
Dyna Robotics' work at Din Tai Fung began as a pilot in 2025. In an August product report, the company described how the initial system could fold a napkin yet still leave restaurant staff with cleanup work. Dyna-1 dropped completed napkins into bins without keeping the stacks organized, requiring employees to straighten them before carrying them into the dining room.
Dyna-2 places each napkin into one of as many as 10 designated positions based on an instruction. Dyna Robotics said it built the placement behavior into the model rather than hard-coding a configuration for each bin. That distinction matters commercially because Din Tai Fung locations use bins with different capacities. A system that requires custom engineering for every container adds installation work each time the robot reaches a new restaurant.
The deployment also forced Dyna Robotics to move evaluation out of the lab. Each production episode is recorded with camera feeds, robot state, control commands, application events and hardware telemetry. Dyna Robotics says its deployed machines generate over a terabyte of raw data each day, which an automatic labeling system breaks into individual steps and failure modes.
That system helped trace a throughput decline at one Din Tai Fung site to missed napkin grabs. Dyna Robotics eventually identified a worn gripper rather than a model regression. The episode captures the less photogenic work behind commercial robotics: distinguishing bad software from aging hardware before either interrupts a restaurant shift.
The company says a new deployment can now move from installation to its production ROI target in as little as three days. Dyna Robotics also expects restaurant, hotel, logistics and data-center installations to bring its fleet into the hundreds during the first half of 2027. Both remain company targets as the Din Tai Fung rollout proceeds.
A second hardware company for Gao and Yang
Gao and Yang previously built Caper AI, whose smart shopping carts used cameras and AI to automate grocery checkout. Instacart acquired Caper AI for $350 million in 2021. Ma, the third Dyna Robotics co-founder, previously worked as a research scientist at DeepMind and has focused on robot learning and foundation models.
That combination explains Dyna Robotics' emphasis on a full product stack. Gao and Yang have already dealt with the manufacturing, installation and customer-support costs attached to AI hardware. Ma brings the model research needed to make one system work across changing objects, sites and instructions.
Dyna Robotics began financing that strategy with a $23.5 million seed round co-led by CRV and First Round Capital in March 2025. Six months later, it raised a $120 million Series A led by RobotStrategy, CRV and First Round, with participation from NVentures, the Amazon Industrial Innovation Fund, Salesforce Ventures, Samsung Next and LG Technology Ventures.
The capital funded a substantially larger training effort. Dyna Robotics introduced Dyna-2 on August 15th as a world-action model pre-trained on more than one million hours of first-person human video. The company says the model can learn robot tasks with limited task-specific data, using human video to improve action prediction before deployment.
Din Tai Fung gives that research a commercial scorecard. The model has to fold fast enough, reject fewer napkins, stack the accepted output and continue operating as materials, lighting and hardware condition change. Dyna Robotics has cleared that bar by its own measurements. The restaurant-network rollout will test whether those measurements hold as the number of machines and locations increases.