Robotics Digital Twin for Kitchen Staff Duties
This one is ours, and it is not finished. We are building a humanoid for the staff duties around a commercial kitchen: clearing a station, wiping down, loading and unloading the wash, putting the room back the way the next service expects it. It learns inside a physics-accurate rebuild of the venue, and it does not enter the building until its behaviour has passed review in that rebuild.
- IndustryFood service and commercial kitchens
- Type of workOwn build, in progress
- StatusRehearsing in simulation, not deployed
What we are building, and what we are not
A commercial kitchen runs on the duties nobody wants at the end of a shift: clearing a station, wiping down, loading and unloading the wash, putting everything back where the next service will look for it. We are building a humanoid for that work. What we treat as the real product is not the robot, it is the rebuild of the venue the robot learns in.
The honest boundary matters more here than anywhere else on this site. Nobody has a robot that cleans a whole kitchen from start to finish, and the field's own open problem is transfer: behaviour proven in one room does not carry to the next one by itself. That is exactly why the rebuild is made for one specific venue, and why review sits between the rehearsal and the door.
Why this is the hardest room to hand to a robot
These are the duties a kitchen struggles hardest to keep staffed, and the room they happen in is the least forgiving one in the building.
The work nobody wants to be rostered for
Clearing and cleaning land at the end of a long service, they are physically hard, and the roles that carry them turn over faster than the rest of the room. That pressure is what makes automation worth attempting here at all.
A room that punishes a wrong move
Wet floors, hot surfaces, blades, glass, and people moving quickly past each other in a narrow space. Anything introduced into that has to behave when it is surprised.
New equipment reopens the hazard analysis
Putting anything new where food is prepared reopens the hygiene and hazard assessment for the whole line. That is a design constraint from the first week, not a certificate chased at the end.
Rehearse in a rebuild of the room, then review
There is no version of this where a robot learns on the floor during service. So the room is rebuilt, the work is rehearsed there, and the gate it has to pass is there too.
The venue rebuilt as a working twin
A walkthrough scan of the real kitchen becomes a physics-accurate version of it, where the things that matter to a body moving through the space are carried over as geometry it can collide with, not just as something that looks right on screen.
- Surfaces, fixtures and clearances carried over as collision geometry
- Doors, drawers and appliance handles reconstructed as moving parts
- Built for the venue it is meant for, and rebuilt when the room changes
Rehearsal at a volume the real room cannot supply
The work is shown by hand first: careful, deliberately smooth demonstrations driven through the robot. Those recordings are then amplified into a far larger body of rehearsal by varying the scene around the same intent.
- Hand-driven demonstrations, kept only when they come out clean
- Systematic variation of the scene to amplify a small set of recordings
- The one venue varied into related layouts, so a room is not memorised
Behaviour that does not depend on one perfect room
The rehearsal deliberately refuses to be comfortable. What the robot sees and when it sees it changes from run to run, so what survives is the task itself.
- Lighting and material appearance varied between runs
- Camera placement and sensor timing changed from run to run
- Where things have been left varied, including the awkward cases
The review it has to pass first
Review is a stage with its own gate. Behaviour is checked against fixed written scenarios as well as randomised runs, so a breaking change shows up as a regression before it becomes a surprise on site.
- Written scenarios alongside randomised runs, checked for regressions
- Safe stopping designed and tested, including which states it can halt from without falling
- Hazardous edge cases generated deliberately
What rehearsing first changes about the risk
None of this is deployed, so what follows is what the method changes about risk rather than results from a floor. It is worth building because how a behaviour performs in the rebuild tracks how it performs in the room, which is what makes a review before the door worth passing.
Before the door
Where mistakes land
In the rebuild a failed attempt costs a run. In the room it costs a wet floor, a dropped tray, or a person in the way. The expensive version of the mistake is the one that never has to happen.
Per venue
How it is meant to scale
Behaviour proven in one room does not carry to the next one by itself, so the rehearsal is rebuilt for each venue instead of being sold as a robot that works anywhere.
Reviewed first
What clears a release
Nothing enters the building on the strength of a good demonstration. It enters when the scenarios pass, the regressions are clean, and the safe-stop behaviour holds.
Considering automation in a room where a mistake is expensive?
The useful first conversation is about the space, the duties, and the constraints they have to respect. Tell us what the room does, and we will be straight about what is rehearsable today and what is not.
Discuss your project