Dyna Robotics' Robots Deployed for Napkin Folding at Din Tai Fung Locations in the U.S.
Dyna Robotics' robots have been scaled for napkin folding at Din Tai Fung locations in the U.S., transitioning from high success rates in the lab to real commercial production, and expanding to hotel laundry scenarios.
Co-founder and Chief Scientist Jason Ma stated that Dyna-1 achieved a 99.4% success rate in 24-hour continuous lab operation, folding over 850 napkins; however, in-store initial performance was only about 35 napkins/hour with a 75% compliance rate, producing around 480 napkins daily, far below the client's requirement of 1,500 napkins in an 18-hour shift.
Dyna-2 increased the speed to 95 napkins/hour with a quality compliance rate of 93%, producing about 1,590 napkins suitable for serving daily, exceeding targets; clients categorize the folded napkins into five levels, only accepting the top two levels, with millimeter and centimeter differences in stitching determining compliance.
The company utilizes over 1 million hours of first-person perspective human video to pre-train the World Action Model (WAM), discovering that more human videos can enhance the robot's action prediction accuracy, even if the model has not seen these specific robots; after early deployment, the faster the robots fold, the more manual replenishment and stacking work increases.
To address this, they introduced the wheeled humanoid robot Taku, capable of autonomously completing the full workflow of washing, drying, folding, and shelving, achieving approximately 79 steps in about an hour of unedited continuous operation during demonstrations.
Jason Ma described the process as "Bittersweet": the difficulty is much greater than expected, and the problems are not glamorous, but the solutions can lead to significant advantages; he criticized the general brain approach research like Physical Intelligence for being strong but lacking successful commercial cases, with downstream clients and upstream models struggling to be profitable.
Market Mechanism: Restaurants and hotels are purchasing and deploying robots to reduce repetitive labor costs, directing funds to hardware + model suppliers that can provide reliable output and quality, benefiting Dyna, which has crossed the ROI threshold and secured multiple contracts, while competitors remain under pressure, either still in the demonstration phase or unable to meet shift output requirements.
Source: Public Information
ABAB AI Insight
Dyna was co-founded by Lindon Gao, York Yang (who was previously at Caper AI acquired by Instacart for $350 million), and former DeepMind/NVIDIA researcher Jason Ma, choosing to focus on fixed dual-arm napkin/towel folding rather than humanoid showcases; the company directly uses deployment feedback for model iteration, advancing from Dyna-1's high lab success rate to Dyna-2's commercial output.
On the capital path, resources are concentrated on collecting millions of hours of first-person human videos and real store data, driving the transition from VLA to WAM, and using the shift output and quality acceptance from clients (like Din Tai Fung and Best Western) as a validation loop; the motivation is to establish ROI with a single narrow task before horizontally expanding to a complete workflow.
This is analogous to early industrial robots on automotive assembly lines performing reliable repetitive tasks or specialized equipment paths in laundry automation; currently, Dyna is transitioning from reliable deployment of single tasks to expanding end-to-end workflows, contrasting with paths focused on general brains but lacking commercial cases.
Structurally, it represents a technological substitution and industrial chain reconstruction: replacing repetitive manual labor with highly reliable narrow-task robots, with the mechanism being defined by customer quality levels and shift output constraints, converting lab success rates into measurable commercial output, making "boring" continuous work a prerequisite for profitability rather than the demonstration itself.
ABAB News · Cognitive Laws
Lab 99% to Store 70%, the gap is in the details
The faster, the busier; workflows are more valuable than single tasks
Boring reliability is when robots truly start working