Industries / Retail, hospitality and personal services / Hospitality and tourism
Restaurant kitchens
Restaurant and catering kitchens cook food to order for each guest, and this work covers how cooks are organized, the ovens and other equipment they use, the rules that keep food safe and the machines that take over parts of the cooking.
6
Milestones achieved, from 1890 to 2018.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
In depth
The same ground is charted in more depth, as an industry of its own.
The road
AchievedWatch forOpen
Achieved
1890
Cooks split into stations
Auguste Escoffier organized cooks into a brigade, a ranked team in which each cook makes one part of the meal: the poissonnier makes only fish dishes and the pâtissier only desserts and pastry. Connexion France says the aim was that "large numbers of meals could be prepared at once", and that by 1890 Escoffier ran the kitchens of the Savoy in London. The Escoffier school places the system's development earlier, in the hotel kitchens Escoffier ran for César Ritz in Monte Carlo and Lucerne from 1884, so sources differ on where it was first used. Kitchens still use the system.
Credit
Auguste Escoffier, working with the hotelier César Ritz
1948
A kitchen built for speed
Richard and Maurice McDonald shut their drive-in at San Bernardino, California, for three months, cut the menu to nine items and gave each worker one task: one grilled the burgers, another added condiments, another made the shakes. They fired their 20 carhops and had customers order at service windows, a setup they called the Speedee Service System, taking a cue from Henry Ford's assembly line. Profits fell at first and by 1951 were 40% above their level before the change, by PBS SoCal's account.
Credit
Richard and Maurice McDonald
1971
Safety checks at each step
HACCP, short for hazard analysis and critical control points, has a food business find the steps where a hazard can be controlled, such as a cooking temperature, and check those steps every time. Pillsbury worked out the method with NASA and US Army laboratories at Natick, Massachusetts, with microbiologist Howard Bauman as its lead scientist, and it was discussed at the 1971 National Conference on Food Protection, by the account of food scientists Richard Stier and William Sperber. Under EU Regulation 852/2004, every food business after the farm, restaurants included, has had to keep procedures based on HACCP since 1 January 2006, with flexibility written in for small businesses.
Credit
Pillsbury, NASA and the US Army laboratories at Natick; the European Union
1974
Cooking in sealed bags
The chef Pierre Troisgros wanted to cook foie gras so that it lost as little fat as possible, and hired the chef Georges Pralus, who found that the liver lost least when poached at a precise temperature. The method, sealing food in plastic and cooking it in water held at an exact temperature, is called sous vide, French for "under vacuum", and America's Test Kitchen dates its arrival in restaurants to 1974. Bruno Goussault made similar findings for commercial food makers and hospitals, and has been chief scientist at Cuisine Solutions, a sous vide food company, since 1989.
Credit
Georges Pralus, for Pierre Troisgros; Bruno Goussault
1976
Steam added to the oven
Siegfried Meister's company Rational, in Landsberg, Germany, made convection ovens, which cook with hot, dry, moving air. In 1976 Meister and a small team added steam to that air, making the combi-steamer, an oven that can cook with dry heat, steam or both. The Craft Guild of Chefs reported in 2017 that Rational's combi-steamers were distributed in more than 100 countries.
Credit
Siegfried Meister and Rational
2018
Robot cooks meet real service
Miso Robotics' burger-flipping robot, Flippy, was switched off at CaliBurger in Pasadena, California, in March 2018, soon after it started, when publicity brought in more orders than it could handle, Fox News reported, citing the BBC. Zume used robot arms to assemble pizzas and ovens in trucks to bake them on the way to customers; cheese slid off on bumps and turns, Zume stopped making pizza in 2020 and was insolvent by June 2023, after raising as much as $445 million. Creator closed its San Francisco robot burger restaurant in 2020 and returned in April 2026 selling its machine to other restaurant brands, leaving what it calls "robot theater". Chipotle ran a brief test of Miso's chip-frying robot in Orange County, California, and called it off in 2024.
LessonA cooking robot has to keep pace with a real rush and fit a working kitchen, and Creator now sells into existing kitchens instead of building restaurants around a robot.
foxnews.comgizmodo.comrestauranttechnologynews.comthespoon.techthespoon.techocbj.com
Credit
Miso Robotics and CaliBurger; Zume; Creator; Chipotle
Watch for
Infinite Kitchen changes hands
Sweetgreen, a salad chain, says its Infinite Kitchen, an automated assembly line for its meals, can "deliver significantly faster throughput and enhanced food quality, accuracy, and consistency". It ran in more than 20 Sweetgreen restaurants when Sweetgreen agreed in November 2025 to sell its maker, Spyce, to Wonder, a food-hall and delivery company, for $186.4 million, keeping the right to go on installing the machines. No measured figures for the machine itself were published.
Achieved whenAn outside party publishes measured throughput or order accuracy from Infinite Kitchen restaurants, or a brand outside Sweetgreen runs one in routine service.
Claimed by
Sweetgreen, November 2025
Fry robot, twice as fast
Miso Robotics' website says its Flippy Fry Station cooks a restaurant's whole fried menu "2x faster than a human" and reports "Zero burns in live deployments". It quotes White Castle staff saying the first units are "doing great" and that the chain has installed a third unit in the St. Louis region. The page gives no method, unit count or dates.
Achieved whenA customer or an outside party publishes measured baskets or orders per hour against a person on the same station, with incident counts.
Claimed by
Miso Robotics, on its website as read in November 2026
Open
01
Check plates against a photo
An open test of camera checks at the pass, the counter where finished plates wait for servers. In a test kitchen, plate five dishes many times, each against a reference photo and a written specification. Build in planned faults on some plates: a missing garnish, a swapped component, a portion 25% light, sauce on the rim. Weigh every component on a kitchen scale before plating. Then have vision-language models, AI models that read images and text, compare each plate with its reference photo, and publish what they catch and what they flag wrongly.
The briefCredit · unclaimed02
Catch cross-contact on video
A labeled video set and a scored test for spotting allergen cross-contact, which is an allergen getting into food that should be free of it, through shared tools, surfaces, oil or hands. In a test kitchen, cooks work through scripted orders, some marked allergen-free, and make planned mistakes: tongs moved from a shrimp tray to the allergen-free plate, gloves not changed after handling nuts, a fryer basket shared, a cutting board not washed. Stand-in foods play the allergens, so no one with an allergy is at risk. Score how many planned events a video model flags, and how many false alarms it raises per hour.
The briefCredit · unclaimed03
Robot scrapes a cold griddle
A robot arm that scrapes and wipes an ordinary flat-top griddle, the flat steel plate on which burgers, eggs and onions are cooked, until it is clean. A person makes the residue by cooking a set recipe on a loose steel griddle plate, then lets it cool and unplugs it. The robot only ever works on the cold, unpowered plate. Photos before and after each run measure the share of the soiled area cleaned.
The briefCredit · unclaimed
Open 01 · the brief
Check plates against a photo
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open test of camera checks at the pass, the counter where finished plates wait for servers. In a test kitchen, plate five dishes many times, each against a reference photo and a written specification. Build in planned faults on some plates: a missing garnish, a swapped component, a portion 25% light, sauce on the rim. Weigh every component on a kitchen scale before plating. Then have vision-language models, AI models that read images and text, compare each plate with its reference photo, and publish what they catch and what they flag wrongly.
Why it is still open
Domino's Australia has run a camera, built with Dragontail Systems, that grades every pizza for type, toppings and how evenly they are spread in all its Australian and New Zealand stores since 2019, but the launch report gives no accuracy figure and this research found none published. A Roboflow tutorial of May 2026 checks plates against a written standard rather than a photo and reports no accuracy on real plates, and Intel's open dine-in checker matches the items on a plate to the order without checking portion or presentation. No public set of plated dishes with weighed components turned up.
Built with what exists
An overhead webcam or phone on a fixed mount with fixed lighting; a kitchen scale reading to 1 g; a vision-language model such as Claude Fable 5.2 (see the Note "Three working devices") and the open MiniCPM-V-4.5 model that Intel's checker uses; an open labeling tool such as CVAT; Roboflow Workflows or a short Python script to run each model.
First step
Pick one dish, photograph a reference plate, then plate it 30 times, 10 of them with planned faults, weighing each component, and run one model on the photos.
Done looks like
At least 500 photographed plates across five dishes, at least 40% of them with one planned fault, each with its component weights, published under an open license, and a table giving each model's catch rate for each kind of fault and its false alarms on correct plates.
Open 02 · the brief
Catch cross-contact on video
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A labeled video set and a scored test for spotting allergen cross-contact, which is an allergen getting into food that should be free of it, through shared tools, surfaces, oil or hands. In a test kitchen, cooks work through scripted orders, some marked allergen-free, and make planned mistakes: tongs moved from a shrimp tray to the allergen-free plate, gloves not changed after handling nuts, a fryer basket shared, a cutting board not washed. Stand-in foods play the allergens, so no one with an allergy is at risk. Score how many planned events a video model flags, and how many false alarms it raises per hour.
Why it is still open
CDC investigators who interviewed staff at 278 US restaurants found that fewer than half of managers had food allergy training and that few restaurants had separate equipment or areas for allergen-free food. Published vision work found in this research, such as Allergen30 (Food Analytical Methods, 2022), identifies foods that may contain allergens, not how allergens move between them. Camera systems sold for allergen changeovers are aimed at factory lines, and this research found no camera system that watches for cross-contact during restaurant cooking.
Built with what exists
Two or three phones or ceiling cameras; stand-in foods with a colored marker; an open labeling tool such as CVAT for timestamps; vision-language models that accept video or sampled frames, such as Claude Fable 5.2 (see the Note "Three working devices"); a public spreadsheet of the script.
First step
Write a script of 20 orders with 10 planned errors, film it in a school or home kitchen from two angles, and label the timestamps of each error.
Done looks like
At least 200 scripted orders containing at least 100 planned cross-contact events of five kinds, with every event timestamped, the video and labels published, and a table giving, for at least two models, the share of events caught by kind and the false alarms per hour of cooking.
Open 03 · the brief
Robot scrapes a cold griddle
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A robot arm that scrapes and wipes an ordinary flat-top griddle, the flat steel plate on which burgers, eggs and onions are cooked, until it is clean. A person makes the residue by cooking a set recipe on a loose steel griddle plate, then lets it cool and unplugs it. The robot only ever works on the cold, unpowered plate. Photos before and after each run measure the share of the soiled area cleaned.
Why it is still open
Makers of automated grills build the scraper into their own machines. Aniai says its Alpha Grill has a robotic spatula that "scrapes the grill surface clean from residue between cycles", and a Taylor Commercial Foodservice patent granted in July 2018 builds a blade into a two-sided grill because otherwise an operator must "manually scrape and clean" the plates, and cleaning them hot "may result in accidental injury to a user". This research found no robot that cleans the ordinary flat-tops most kitchens already own; the kitchen robots found cook food, such as Miso's fry station, or wash dishes.
Built with what exists
A low-cost SO-101 arm with Hugging Face's LeRobot v0.7.0 software for a first attempt (see the Note "No download"), or a collaborative arm such as the Universal Robots UR5e for more force; a commercial griddle scraper and scouring pad; a color camera with OpenCV to measure residue area; a robot model that learns from a few demonstrations, such as NVIDIA GR00T 2 once its weights can be downloaded (see the Note "Five demonstrations per task"). Robot models that learn a task from a few demonstrations are what bring this within reach of a small team.
First step
Buy a steel griddle plate, fix a residue recipe (for example, the same weight of onions and sugar cooked for ten minutes), let it cool, photograph it, and record five demonstrations of scraping it with the arm's leader arm.
Done looks like
Twenty filmed, uncut runs on plates soiled by the same recipe, with a public table of the share of residue area removed, minutes per run and human touches for each run, and a person cleaning the same plate five times for comparison.