Concept visual: Wild Bite Club.
One high-volume linen task reveals why adaptable automation needs instruction-following, fleet telemetry and relentless work after the demo.
How It Works
A cloth napkin is an awkward object for conventional automation. It changes shape, collapses under its own weight, arrives with wrinkles and must be grasped, aligned, flattened, folded and placed without a rigid fixture controlling every move. A system that works only when the fabric and bin are positioned exactly like the lab demonstration will fail as soon as a station is rearranged.
The technical answer is a general manipulation model rather than a machine programmed for one fold. Dyna-2 is described as a world-action model pre-trained on more than one million hours of first-person human video. That video teaches the model how objects and hands move through physical interactions; a smaller amount of task-specific robot data then adapts the behaviour to a particular machine and workflow.
Instruction-following closes another operational gap. A finished napkin must land in an organised stack, not merely somewhere inside a container. Different sites use bins with different capacities, so a command that specifies one of several placement positions is more reusable than a separate hard-coded routine for every bin. The fleet then supplies a feedback loop. Camera streams, joint state, control commands and hardware telemetry are segmented against the restaurant’s standard operating procedure. Repeated failures can be traced to a step, site or component. In one company example, rising missed grasps led the team to a worn gripper rather than a bad model. The mechanism is therefore not one clever arm. It is adaptable control combined with continuous production observability.
What Is Changing
Restaurant automation has often been sold through visible theatre: a robotic arm flips food, pours a drink or carries a plate. The more instructive experiment may be happening away from the recipe. A commercial rollout at Din Tai Fung is built around folding and stacking table napkins, a repetitive support task with a clear daily requirement and an unforgiving quality standard.
Dyna Robotics says each machine must supply about 1,500 table-ready napkins during an 18-hour shift. Its earlier system produced roughly 35 folds an hour, with three quarters clearing the restaurant’s quality bar. The newer system reportedly reaches 95 an hour and sends 93 percent of the output forward, yielding about 1,590 acceptable napkins a day. Those figures are company-reported rather than independently audited, but the decision to expand beyond pilot sites provides an unusually concrete operating signal.
Why It Matters
The choice of napkins matters because it reverses the usual restaurant-robot pitch. The machine is not trying to replace the craft that defines the brand. It is taking on a high-volume task around the meal: necessary, measurable and largely invisible to guests. That lowers culinary risk while still addressing labour availability, consistency and repetitive strain.
For consultants, this creates a practical automation sequence. Start with a task whose inputs and outputs can be defined, whose failures are easy to inspect and whose economics can be counted per shift. Only then consider moving closer to food. The restaurant also retains a human-controlled culinary core. That separation may be especially important for premium or culturally specific concepts that want operational leverage without turning their signature dishes into a robotics demonstration.
Where It Could Go
If that operating stack proves durable, foodservice automation could expand task by task rather than arrive as a fully autonomous kitchen. The next opportunities may be jobs with similar characteristics: sorting clean serviceware, preparing amenity kits, moving packaged ingredients, loading trays, clearing workspaces or performing simple, low-risk assembly.
Movement toward food would raise the difficulty sharply. Ingredients are wet, deformable, temperature-sensitive and variable; sanitation and allergen controls introduce constraints that napkins do not. The most plausible path is therefore conditional. A general-purpose platform might first earn trust on support work, collect months of site data and establish service economics. It could then approach selected food-contact tasks where the product, station and cleaning protocol are sufficiently controlled. In that model, the robot does not begin as a chef. It becomes a flexible piece of operating equipment with a growing library of validated jobs.
What to Watch
The next evidence should come from operations, not demonstrations. Watch for the number of restaurant sites, the exact task at each location, paid contract terms, uptime over full shifts and the share of output that needs human correction. Independent confirmation from the restaurant operator would be more valuable than another supplier benchmark.
Maintenance will be decisive. A robot that folds quickly but requires frequent intervention can move labour rather than remove it. Service response, gripper wear, changeover time and recovery after an error should be measured alongside throughput. Consultants should also ask how much site-specific training is required and whether improvements learned at one restaurant transfer to another. Before any food-contact work qualifies, the evidence bar should add cleanability, food-safe materials, temperature control, allergen separation and performance with irregular ingredients. The real milestone is not a broader list of possible tasks; it is a repeatable deployment playbook with predictable economics.
The WBC Read
This is a strong foodservice experiment because the innovation sits in the deployment logic as much as in the robot. Choosing napkin folding creates a bounded, commercially meaningful proving ground, while instruction-following and fleet telemetry address the unglamorous failures that usually separate a demo from daily work. The restaurant rollout makes the proposition more credible than a laboratory video alone.
The caution is that most detailed performance figures come from the supplier, and success with linen does not establish readiness for cooking. Fabric handling does, however, test adaptability, precision and sustained operation in a real restaurant environment. The case is worth following as a model for automation sequencing: protect the food experience, automate a measurable support task, and let production evidence decide what the machine earns the right to do next.