AI Engineer World's Fair 2025
General purpose robots as professional Chefs
Loading slides…
Read the talk
General purpose robots as professional Chefs
CloudChef combines robot motion skills with thermal and visual perception to reproduce a chef’s cooking, while teleoperators and kitchen staff handle the remaining gaps.
From a talk by Nikhil Abraham
Two arms, a mobile base, and a kitchen job
What does it take to turn a general-purpose robot with two hands into a professional cook? For CloudChef, the goal is affordable, nutritious, high-quality food through automation of commercial kitchen labor. Nikhil Abraham, its co-founder and CEO, starts with a practical hardware choice: a robot must act, sense, and reason in a kitchen, but it does not necessarily need a humanoid body.
Abraham contrasts humanoids’ cost, maintenance demands, and reliability with two arms mounted on a wheeled base. He presents the latter as already cheaper than human kitchen labor, leaving software as the missing ingredient. The opening cost chart places Zippy 1.0 near the point where the falling robot-cost curve sits below the rising human-cook-cost line.
Abraham quotes robot labor at about $12 per hour in the talk. The commercial proposition includes showing up without turnover or overtime and working in an unfamiliar kitchen. He describes learning a recipe from one expert demonstration, adapting to ingredient and appliance variation, and cooking different portion sizes. These are the capabilities the subsequent demonstrations set out to explain.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Motion skills are only the beginning
The first part of culinary school is physical: picking something up, stirring a pot, and performing the other motion primitives that people bring to a kitchen. CloudChef fine-tunes robot foundation models for these skills and falls back to teleoperation for edge cases.
The next part is knowing what the motion has accomplished. Are the onions brown enough? Has the shrimp shrunk enough to indicate doneness? Ingredients vary from day to day and season to season. Abraham’s illustrative example is onions that need seven minutes of sautéing one day and nine the next. Replaying yesterday’s duration would not reliably reproduce yesterday’s result.
A recipe is represented as a state machine, with cooking-specific thermal and visual embeddings at its core. Those representations help the system reason about food in environments it has not seen before. The remaining challenge is to bring that culinary understanding into a new kitchen, where the robot must identify what to do and fit into an existing workflow.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Measuring cooking progress separately from movement
To test culinary understanding, CloudChef asks a narrower question than whether a robot is a better chef overall: given live cooking data, an expert demonstration, and a text recipe, can it identify where the dish is in the cooking process? Abraham reports an evaluation covering more than 1,000 recipes across a mix of cuisines.
On this cooking-progress task, Abraham reports that CloudChef’s small perception model outperformed expert chefs, including chefs earning more than $150,000 annually. He also reports that Gemini 2.5 and o3 performed worse, attributing part of the difference to their lack of a thermal modality in this comparison. The talk supplies neither numerical accuracy nor a complete protocol establishing the recipe split, model configurations, or input parity. The comparison therefore concerns the reported state-estimation task, not general cooking ability.
Thermal sensing creates a data problem: there is no corresponding supply of internet-scale thermal cooking data. Abraham says CloudChef collected sensor data from hundreds of thousands of meals cooked in active commercial kitchens. That collection spans kitchens, recipes, cuisines, and seasons. The training strategy combines this private data with scraped public data and self-supervised models.
Motor performance is a separate limitation. Sautéing is described as almost as fast as a human cook, while picking and pouring remain slower; grilling and stirring also appear among the evaluated skills. Abraham reports operation that is about 95% autonomous and 5% teleoperated, without specifying the denominator or observation period. He describes the hybrid system as faster and more reliable than either teleoperation alone or foundation models alone.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Match the onions, then monitor the chicken
The recipe demonstration makes the state-estimation mechanism concrete. Abraham attributes the recipe to a San Francisco chef with two Michelin stars. As the robot cooks, it compares the onions’ current browning with the browning observed during the chef’s demonstration, bringing them to the corresponding state. The displayed pan view outlines its contents in red and labels the operation “Tracking Onion browning.”
Abraham says the system reasons about where ingredients are kept and what variations it encounters, rather than relying on preprogrammed ingredient locations. The demonstration then progresses to chicken, with repeated readings described as occurring every few minutes. Those observations illustrate ongoing monitoring; they do not specify a food-safety measurement protocol.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From a recipe demonstration to customer meals
The presentation moves from reproducing a recipe to footage of deployed work. Abraham says the meals reach actual customers and distinguishes footage from CloudChef’s in-house kitchen, CCTV footage of robot operation, and a chicken-wing workflow at a customer facility.
The wing workflow follows a familiar kitchen sequence:
- Fetch the wings from their staging location.
- Wait for cooking to finish.
- Transfer the cooked chicken into a bowl.
- Add sauce and mix.
Abraham describes the robot as a weighing scale that knows how much ingredient it has added, while also tracking how much it has stirred. He closes the prepared presentation with a recruiting invitation for a small team working across software, machine learning, and robotics.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Progress estimates, motion speed, and blind tasting
The audience first asks what it means to perform almost like a human. Abraham separates two measurements. For perception, a person interprets cooking footage while the system interprets video and infrared feeds. A completed cooking sequence then provides retrospective supervision for intermediate estimates: if the model predicted 40% done, was the corresponding state actually 40% or 50%? Recorded recipe recreations with thermal and RGB footage supply that learning signal.
For motion, the comparison is how quickly the robot performs a physical task relative to a person. Abraham acknowledges that the robot still trails humans and attributes further improvement to collecting more data for individual kitchen tasks. This keeps perception quality and movement speed from collapsing into a single ambiguous measure of success.
A follow-up asks about the meal itself. The target is consistency relative to a chef making the dish again, rather than chemical identity between servings. CloudChef’s in-house procedure is to have a chef cook a reference recipe, have the robot recreate it several times, and conduct blind taste tests. Abraham describes these tests as less scalable but a stronger signal about the final product, and reports favorable results. No sample sizes, scoring details, or uncertainty estimates accompany that claim.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
What arrives in the kitchen
The delivered form factor is two arms on a mobile base, with cameras and related sensing equipment. The screens shown in the video are presentation aids, not additional equipment included with the robot. Abraham describes interacting with the unit as one would with a human coworker.
Ingredient measurement is not entirely hands-off in every deployment: humans still prepare measurements in some kitchens. The robot’s own weighing capability comes from joint torque data across its motors. Abraham explains that when it lifts an object, those readings let it estimate how heavy the object is.
For appliances, the intended combination is sensing, reusable motion primitives, and control systems. Abraham says the system can work with previously unseen appliances; because many kitchen appliances use knobs, turning a knob is a particularly useful primitive. The control system then takes over regulation. The wheeled base also supports automatic movement around the kitchen.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Why motion data limits speed
Asked how quickly the robots can move, Abraham reports that most motions run at 80–95% of human speed. His explanation is that people supply the training demonstrations by teleoperating robots, and controlling a robot is less intuitive than moving one’s own body. The resulting demonstrations are slow, so the training data carries that limitation forward. Reinforcement learning is a proposed route to faster behavior, not a measured improvement presented here.
Chopping and dicing are absent from the motion comparison because the team has not yet collected the necessary task data. They remain on the roadmap rather than among the demonstrated capabilities.
For people wanting to try the food, Abraham names Wingstop in a San Francisco ordering example, India’s Top 20 by CloudChef in Palo Alto, and a high-end Indian restaurant in Menlo Park that he calls Elan. He describes some food at the latter as robot-cooked. These are deployment examples offered at the time of the talk, not a promise that every branch, dish, or current order involves a robot.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Kitchen throughput is more than arm speed
The next question shifts from individual motions to service throughput: how much do preparation, slower movement, and backup teleoperators matter to the economics of a working kitchen? Abraham estimates that line cooking accounts for about 50% of kitchen labor costs, making it CloudChef’s initial target. His answer focuses on that labor category and recipe duration, without quantifying teleoperator costs or total kitchen throughput.
A robot can potentially compensate for slower movements by shortening the cooking process itself. Abraham reports instances in which a chef’s 20-minute process could be completed in 14 minutes. He uses a robot moving 10% slower as an illustration of why modest motion disadvantages need not determine overall performance. These are process examples, not a general recipe-speed guarantee.
Longer operating hours offer another potential advantage. Abraham contrasts a human entering overtime after 40 hours with a theoretical 168-hour robot workweek. That is a possibility of continuous operation, not a demonstrated uptime result: most facilities do not run around the clock, so their schedules constrain actual use. Overnight cutting and chopping are envisioned only after those capabilities are implemented.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The work that still needs people
Questions about dishwashing and cross-contamination expose the boundaries of the initial product. Abraham describes small washable silicone pads or gloves on the robot’s contact surfaces, which customers replace daily. That is the handling practice described in the talk, not a comprehensive sanitation validation. Dishwashing is outside the near-term scope; the priority remains line cooking and tasks that directly affect food quality, with preparation and chopping possible later additions.
Asked directly whether the robot can modify recipe steps to cook faster, Abraham qualifies the earlier speedup examples: recipe optimization remains experimental and varies by cuisine. It is easier where the cooking process lends itself to straightforward thermodynamic modeling, which can support shorter cooking times and minor variations. The team has not extended that capability across all cuisines.
The final question is what happens when an ingredient such as sauce runs out. In the current system, the robot alerts someone in the facility, and that person handles replenishment. Coordination among multiple robots is a future aspiration. For now, the robot can perform cooking work inside a human-run kitchen, but the kitchen staff still keep that work supplied.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
CloudChef's product overview, robot form factors, recipe-learning claims, and current hourly pricing.
The Palo Alto kitchen's menu and pickup or delivery ordering page.
Further reading
A pre-conference interview covering culinary intelligence, demonstration learning, commercial applications, and pricing.
Full transcript
Loading transcript…