Scooping Food Into Boxes Is 70 Percent of Plant Labor

Nathan Labenz introduces Rajat Bhageria, founder and CEO of Chef Robotics, as a rare guest: not another robotics-research episode, but a product deployed at non-trivial scale in the wild. Chef builds AI-enabled robots for the food industry and is starting in manufacturing, where it can collect training data while returning ROI — L2 autonomy, in self-driving parlance. More robots mean a better dataset, which means a more dexterous system, which is supposed to unlock ghost kitchens and fast-casual lines later. The endgame is an AI-enabled robot in every commercial kitchen.

The surprise in the intro is the labor split. Food assembly, Labenz was startled to learn, is an estimated 70 percent of labor cost in manufacturing facilities, more than ingredient prep and cooking combined. Chef’s bots have assembled some 20 million meals. Looking ahead, Bhageria sees commercial kitchens of every kind, then a reorganization of the industry: cheaper, more customizable meals from robotic ghost kitchens, delivered by small autonomous vehicles. Against the Jetsons picture of a domestic cook, he expects high-quality food to approach the cost of ingredients, and home kitchens to become an optional add-on for hobbyists.

What “food manufacturing” conjures for most people is ketchup already running on a dedicated Kraft Heinz line. That is low-mix: ten SKUs, ten lines, all day. Meals are high-mix. Prepared sandwiches, salads, party trays, and frozen dinners at Costco or Trader Joe’s, hospital trays, school lunches — vegetarian, vegan, gluten-free, meat-lover, each with roughly 20 SKUs underneath — leave a plant with hundreds of SKUs. Status quo is a long conveyor, people scooping from big tubs into individual containers. That scoop is the wedge Chef is going after first.

The hardware is deliberately dull. Modules sit on casters — a mobile manipulator that customers move maybe once or twice a day, so it does not need to be autonomously mobile. A six-axis collaborative arm, RGB-D cameras, force-torque sensors, and scales do the sensing. Different utensils go on a pneumatic end effector. The module is the same footprint as a person, so it slides onto an existing line with no retrofit. Mass-manufacture one design, mostly off-the-shelf parts, and put the blood, sweat, and tears into autonomy.

Every commercial kitchen has three processes: prep, cooking, and assembly or plating — the Chipotle move of scooping from a hotel pan into a burrito, bowl, or tray. Except in fine dining, prep and cooking scale sublinearly: more volume does not require linearly more humans. Assembly does. Food trucks, fast casual, prisons, hotels, universities, K-12, airline catering, manufacturing — the labor pile is “from tub into box.” Labenz’s picture of a robot that chops onions all day is the wrong station. One arm per module for now, multiple utensils if needed. Just doing assembly is already a gigantic market.

Walk a production room and you see twenty people on either side of a conveyor, each with a tub, a utensil or their hands, scooping into trays, wraps, burritos, sandwiches. The room is often 34°F. Workers touch frozen ingredients and lose feeling in their fingers after five minutes. Turnover is extreme; the job is hard to hire for. Chef’s pitch for this go-to-market is a human equivalent that lets the plant move people onto higher-value work: continuous improvement, material handling, the jobs humans should be doing. The 70 percent stat reminds Labenz of a coding-tools founder who said developers actually write code maybe a third of the time. The labor hog hides where intuition does not look.

Robots ship in wooden crates, already on wheels. Slide onto the line, wash down, take a bacteria swab so the plant is not introducing pathogens. Two facility inputs, both standard: 110-volt AC and compressed air for the pneumatic tool. The day-of user is often not technical, not a four-year-degree holder, and not even English-speaking, so the interface is meant to be as simple as a home oven. Gloves on a touchscreen: today’s meal is pad thai, robot one does the noodles. Attach the recommended utensil, load the tray, and the system fine-tunes itself — pick and dump from one pan to another — before anyone hits play. The UI runs in Spanish or whichever language the floor needs.

That on-shift tune is not a nicety. Food is organic, cut and cooked by humans, and every person, every day, every batch is different. The cook went longer or shorter; the curry is denser in one part of the pan; the noodles are stickier or oilier than yesterday. There is no such thing as two identical batches. Fine-tune, press play, start placing onto the line. First deployments, Chef’s solutions engineers do the ingredient onboarding so the customer is ready on day one. After that, software is supposed to let a production or planning manager self-onboard the meal running next week, or tomorrow.

Customers already live on “meal cards,” PDFs that say noodles, here is an image, this portion, into compartment two, seven or eight ingredients to a card. Upload the card to a web app. A vision-language model parses the unstructured metadata into Chef’s database. Name and portion size alone will run the robot, but consistency will be poor — spills, sloppy placement. So an LLM asks a few questions with generated comparison images and a slider: how long are the particles, from a bite-size piece to a long noodle; how sticky, from cream cheese to water. A minute of that, plus the image, string name, portion, and compartment, plus thousands of past ingredients and (at the time of the taping) some 18 million production servings, yields a starting policy. Deploy it, slide in this customer’s pan, and fine-tune again to *this* batch of *their* cut and cook.

Labenz hears a foundation model for the form factor, with weights actually changing when the noodles miss the spot. Bhageria splits it. Chef was in production a couple of years before foundation models were a thing, so some pieces are CNNs and older deep nets; they are now generalizing a food-manipulation model, then fine-tuning on the cell. Closed-loop metrics decide whether the policy is converging: pick weight versus target (bias of the mean, and spread), placement QA for spread and spill into the wrong compartment, and reliability proxies such as force-torque paths that avoid 150-newton protective stops — a 30-second line pause that is not the end of the world and also not great — plus pressure and flow so the tool does not crush the food. Pick weight is the metric customers care about most.

People expect the robot to be perfect, the way they expect a robotaxi to be perfect, and they assume “AI” means flawless. The food is not. Solutions engineering therefore takes human baselines as blindly as it can on first deploy. Customers often do not believe how bad their own people are against the tight bounds in their heads. That baseline becomes the bar Chef has to meet or beat. FDA, USDA, and eaters punish under-portioning with fines and churn, so leadership tells line managers always to over-pick. A 1-ounce target commonly leaves the belt at 1.2. In food, 40 to 50 percent of COGS is the material; the industry calls the extra “giveaway.” Closing that loop is direct margin. Millions of robots across manufacturing, ghost kitchens, fast casual, prisons, and hotels would also be a macro story about waste.

Numbers depend on customer and ingredient. Broccoli pieces are large — four or five to a scoop — so one floret too many or too few blows the weight; humans might run 25 to 30 percent bias, Chef about 5 percent, still biased up to avoid the under-pick. Mashed potato, sauce, diced vegetables: humans around 4 to 5 percent over the mean, Chef about 1 percent, guaranteed lower because of the control loop. For spread they use coefficient of variation, standard deviation over target, so big portions do not look worse by default. Humans land around 16 to 18 percent CV; Chef’s best is 5 to 6 percent, average more like 7 to 8.

Drop-in and “human equivalent” do not mean slower. Customer lines often run 15 to 20 trays a minute; veterans pick a mass and deposit several times to amortize transport. Chef is around 20 trips a minute, with a multi-utensil system when someone needs more. Why not an industrial arm? Philosophy: partial automation. Full-line autonomy has to handle every SKU or the line still needs a human — long sales cycles, risky deployments, an impossible technology bar. If Chef cannot do shrimp, a person does shrimp. Sales and deploy stay short, the flywheel turns, and the robot always stands between humans, so cobots and their force and velocity limits are a requirement, not a preference. Three or four seconds a trip is about a thousand servings an hour; 20 million servings is about 20,000 hours, on the order of ten person-years, across six cities and five states in the U.S. and Canada since mid-2022. Revenue went 4x from 2022 to 2023, with something like 3x expected this year. A plant does not hire five scoopers; it hires 50 or 100, and whales have multiple plants. A hundred million dollars of ARR, Bhageria says, is 20 to 30 customers if customer success holds.

Foundation models, Bhageria thinks, have uncapped a lot of robotics potential, because the world is high-dimensional: picking up a wine glass is not picking up a teacup. Every customer cuts and cooks differently, every shift, every country. VLMs and LLMs are good at high-level scene understanding because they trained on internet image-language pairs. There is no motion data. Chef cannot download the internet as a food-manipulation corpus; the corpus does not exist. Simulation helps self-driving some of the way, and rigid palletizing can live in a physics engine. Food is wet, sticky, deformable, sometimes non-Newtonian. Grasping is the hard part, and grasping is what you cannot sim. The 20 million servings matter because production is where the training set has to come from.

Where the new stack helps is transformers for imitation learning and transferring knowledge across ingredients: encode one customer’s sticky rice and retrieve a cousin at another plant that cooked it differently. A Google-led, many-lab arm dataset might help the web-app onboarding step; Bhageria has not tried it and will not fake a yes. Train on rigid bodies and you will not manipulate malleable ones, and food demo videos rarely do the primary handling of dripping, sticky, compressible raw material you cannot uncompress. The in-house bet is imitation learning, reinforcement learning, and real production data toward a generalized food-manipulation model, starting at assembly because that is most of the labor, later maybe sliding baker’s trays into a rack.

Labenz asks about NVIDIA’s Eureka, where GPT-4 wrote reward functions for a robot hand better than the humans in the loop. Bhageria translates to parameterized versus non-parameterized policy. In 2019 the precedent was AV-style rules plus CNNs: a policy per ingredient, about 200 human-set parameters — pressure, flow, dwell, number and amplitude of shakes — plus a control loop on pick weight, spillage, placement, reliability, and no idea whether you were at a local or global maximum. Imitation learning puts two identical robots side by side. One mimics a human scoop, same embodiment as the worker on the line, so there is no grounding mismatch, and the policy is end to end: no human-written parameters, not a bajillion lines of code, no cinematic Eureka, and it works. Time-to-onboard a new ingredient is the metric that matters, because coverage is utilization, utilization is more robots, more robots is more data. Outer flywheels — case studies, bill of materials, a denser field-support map — all hang on ingredient coverage.

Each module has two computers: a safety-rated robot controller for joints and motor voltages, third-party certified, and Chef’s own stack for high-level task planning, trajectory and placement strategy, and a loop that talks to the arm controller. The scene is not a robotaxi, but it is hectic. A six-hour run goes through hundreds of hotel pans, each cooked by an imperfect human; pulled meat varies in density inside a single pan. Humans nest trays at odd angles and distances. Conveyors differ in color, width, and height. Floors slope for drainage; the belt itself is a long, vaguely sinusoidal plane, shaking, as is the robot. Someone bumps the line and it skews, so you have to place farther and higher. A line manager cranks speed in the morning. Neighbors pull and load pans and change the scene constantly. Scooping in a lab is easy. Twenty million servings is the work of surviving those edge cases on dozens of systems around the country.

Training runs in the cloud on A100s; inference is all on-device, because plant Wi-Fi is spotty and LTE often dies inside the building. The arm is about 200 watts, the CPU about 35, the GPU about 50. Early customers asked whether the electricity bill would explode; the math is on the order of five dollars a day, even running around the clock. Labenz, in a Detroit market just under 20 cents a kilowatt-hour, notes that 300 watts is a couple of dollars for 24 hours, and treats that as evidence that AI energy panic is overstated relative to the adoption those same leaders imply.

They do not teleoperate the way people imagine. Production customers have no patience: it works, or they pull the robot and put a person in. Worst case is inconsistency, a missed pick, or a spill — not an injury. Four layers handle the rest. First, autonomy: adapt to material, tray pose, conveyor, the human who snatches a tray mid-track. Second, a 3D view of what the robot thinks it sees, so a manager can *see* that the belt is outside the operating domain or the trays are out of reach, without a paragraph of error text. Third, a few coarse knobs for pick and place, few enough that an operator cannot make it much worse. Fourth, telemetry that self-annotates events — huge over-pick, miss, double deposit, pick hesitation — packs 20 seconds of ROS bag on either side, and pages Slack by day or PagerDuty by night, escalating from customer support to applications to engineering. Imitation between robots still happens at Chef’s office, aimed at new, hard ingredients, not as a customer chore.

The frontier is still food manipulation. Nobody else has a serious training set; sim will not save them. The robot can be handed more or less any ingredient and move it. Consistency is the gap: a soft item gets damaged, sticks to the tool, spills, misses the compartment. Flexibility usually costs reliability, and Chef needs both, plus speed, plus any conveyor and any container. Labenz asks whether Tesla, Figure, and the other humanoid shops are “good luck, food is last on the list.” Bhageria says yes. Walking around a production line or a ghost kitchen adds almost no value. The station is the station; wheels move the module once or twice a day. A humanoid spends half its compute not falling down and inherits a new safety case. Rip out a prep table, slide the module in. The 2050 picture he wants looks a bit like *WALL-E*: tens of millions, hopefully hundreds of millions, of robots that are very good at one thing, the way a refrigerator does not do laundry, plus a smaller fleet of generalists for hospitals and homes with fifty unlike tasks. Uber was a business; Uber ATG was always ten years out. Humanoids may take twenty, as AVs took twenty or thirty and are still not done. Software was hardware-agnostic from day one, so if he is wrong, the food model can ride someone else’s body. He does not think it makes sense to hang car doors with a humanoid. The market in front of Chef is large enough to execute on now.

The commercial model is robotics-as-a-service: a yearly fee always below the cost of the status-quo workers. Customers avoid a large upfront outlay, take it from payroll rather than capex, and skip a capital committee. Chef carries GPU, cloud, and the cost of constantly upgrading hardware and software. No capex sales. Layer one of the vision is still manufacturing: an industry short on people, long on demand, and a chance to cut waste. Bhageria thinks the bigger impact is daily life. Early on he gave friends two options. A: a Jetsons robot at home that does whatever you want. B: food made by a robot in a ghost kitchen — cheaper, consistent, nobody sneezes on the guacamole, fully customizable — in cheaper real estate, because about a third of restaurant or delivered-food cost is rent, then delivered faster and cheaper than a human driver. Everyone picked B.

They wanted to keep brands like Sweetgreen. They did not want to maintain a machine that would break. They would rather rent the meal than buy the robot, the Airbnb logic. They did not want to refill groceries. Cleaning, he grants, you cannot outsource from the house. Textiles used to be a cottage industry; factories ended that. He argues food follows, except the factory is a ghost kitchen. An interface lets you specify salt and the rest; a robot cooks it; a robot brings it; the price approaches cooking at home. Studies he cites say 80 to 90 percent of Americans do not like to cook. Kitchenless apartments, and a lot more robotic delivery, are the tailwinds he is willing to name. Labenz calls kitchenless housing one of the least guessed pieces of an AI future, and a reminder that enabling tech can reorganize daily life away from the naive first picture.

Zoom out once more and the claim is that AI’s largest impact is the physical world. Software is wonderful and, relative to global GDP, small. The labor market is about $50 trillion. Food is the third-largest U.S. employer by headcount; retail clerks and nursing aides, one and two, are more human and harder to automate. Tens of millions of food workers will feel this wave early. If Chef becomes this era’s Facebook against Friendster and MySpace, or Intel against the other chip startups, Bhageria wants the legacy to be that the next generation of engineers, founders, and investors does AI plus robots, not only pure software. He closes by asking more people into embodied AI, toward the Jetsons future — even if his own bet is that the butler never moves in.