Chef Robotics

We're heading to IDDBA 2026 in Orlando next week! 🎉 Food manufacturing is changing fast, and we'd love to show you what's possible with physical AI on the production line. Stop by booth 4961 between June 7–9. Whether you want to see a live Chef robot demo, talk automation, or just say hi—we'll be there. See you in Orlando! 👋 #iddba #foodautomation #foodmanufacturing

When Chef robots pick food from a pan, the pan changes constantly—ingredients shift, deplete unevenly, and clump differently depending on their type. A good pick location at the start of a shift is often a bad one just 15 minutes later. Picking well means optimizing for many things at once: consistent volume across tools, collision avoidance, and cycle-time efficiency, among others. Our pick-scoring pipeline meets these competing goals by composing a stack of independently tunable subscorers. Each one is a focused heuristic or physical AI model output that produces its own 2D heatmap over the pan. These subscorers are dynamic and adapt as the module runs. It leverages live perception data, on-device failure memory, and the system's own pick history to converge towards more consistent results per shift. The image below shows what this looks like in practice: a pan of granola on the left and, on the right, a grid of heatmaps showing how different subscorers are scoring the same pan at the same moment. Red means a high score; blue means a low score. Each frame reflects a different signal that the system weighs before committing to a pick location. The pick-scoring pipeline is what keeps pick quality consistent from the first scoop to the last, and our robots don't degrade as a pan empties. They adapt.

Will physical AI be general-purpose or specialized? We think it's both. There's a major debate in robotics right now about whether generalist models and humanoids will outperform verticalized solutions. Our view is that the industry is converging toward a different outcome: 🧠 Intelligence will become increasingly shared 🤖 Embodiments will remain highly specialized Why? Because industrial customers don't buy versatility. They buy ROI. A food manufacturer doesn't need a robot that can fold laundry. They need a robot that can portion food accurately, operate in 32°F cold rooms, survive daily washdowns with caustic chemicals, and run reliably at production scale. The same logic applies across construction, logistics, agriculture, manufacturing, and healthcare. Different industries have different environments, economics, and performance requirements. One embodiment won't fit them all. In our latest blog, we make the case that the future of physical AI will consist of shared intelligence layers combined with verticalized embodiments, proprietary data, and domain expertise. Or put differently: the future may look a lot more like 𝘞𝘈𝘓𝘓-𝘌 than 𝘐, 𝘙𝘰𝘣𝘰𝘵. Read the full post: https://lnkd.in/e_q72yiK #PhysicalAI #Robotics #AI

We’re excited to welcome Steve Van Der Hoeven to Chef as a Senior Staff Software Engineer! Steve brings over two decades of software engineering experience from companies like Google and Optimizely. He most recently worked at Zeromatter, where he built a CI/CD and validation platform for robotics stacks. At Chef, Steve will help our growing team of robotics and software engineers bring robots into real production environments across over a dozen customer sites. If this sounds interesting to you, see our open roles at https://lnkd.in/g2J6g63b! #robotics #ai #hiring

When a robot on a food production line deposits an ingredient into a tray, it doesn't act alone. The conveyor has to stop, every robot on the line has to finish its deposit, and only then does the line move again. That coordination happens dozens of times a minute, every shift. Chef has a device that manages this entire sequence. And until recently, the only way to test whether that logic was working correctly was to run it on a live production line with physical hardware present. We had two ways to test a stop-and-go conveyor without hardware. Neither covered the full loop. So we built a hardware abstraction layer within the runner and an in-process PLC model to simulate the customer controller. The complete indexing stack now runs on a developer laptop with zero hardware in the loop. Read more on our engineering blog: https://lnkd.in/gaSVVMhB #physicalai #robotics #foodrobotics #conveyortypes

Last week, we were making burgers. This week, we’re scooping burrito bowls! Our latest engineering blog explores how we taught Chef’s Food Foundation Model (FFM) to manipulate loose, deformable foods like rice, beans, lettuce, cheese, and chicken using the same underlying physical AI architecture. Scooping sounds simple, but it introduces entirely new robotics challenges: • Portioning loose ingredients precisely • Preventing spills and cross-contamination • Handling food that behaves differently with every scoop Instead of rebuilding the system for a new meal, we trained the same model on new demonstration data and taught our robot to use ordinary kitchen utensils with a robot-friendly handle. After about 25 hours of demonstrations, our robot can assemble a burrito bowl in under 2 minutes. This is what general-purpose physical AI for food looks like. Read the full blog: https://lnkd.in/g85AqD-a #physicalai #robotics #food

An experienced food assembly line worker doesn't pick ingredients from a bin the same way on day one as they do after six months. Our robots now do the same thing. We built a 6D sparse failure maps feature for reactive pick learning. This is an on-device experience model that records every failed ingredient pick from the bin across the full 6-DoF action space—position, depth, approach angle, and gripper twist. The model updates continuously as the robot operates, adapting in real time to the specific characteristics of every site, system, and SKU it runs on. Before each pick, the planner queries this learned memory and steers the robot away from pose regions that have already proven unreliable on that line, ensuring the robot picks from the best possible spot every time. In the visualization below, the red markers represent failed pick poses that the robot has logged. The blue arrow indicates where the planner sends the robot instead, towards the most reliable pick available. As a result, the physical AI model becomes more reliable the longer it runs on a given line. That's not just because of fleet-wide training, but because it's building a failure map that makes each ingredient pick more precise.

The best stories are often the unexpected ones. Robots preparing medically tailored meals in San Francisco’s Tenderloin is one of them. We’re proud to collaborate with Project Open Hand to help assemble medically tailored meals for seniors and community members living with chronic illnesses. For over 40 years, the nonprofit has built its mission around the idea that food is medicine, recognizing the critical role nutrition plays in managing chronic illness. But since the COVID-19 pandemic, the organization has faced ongoing volunteer shortages. That’s where our robots come in. Today, two Chef robots work alongside volunteers in the Project Open Hand kitchen, helping assemble meals and freeing up volunteers to focus on other, less repetitive tasks. This is just one example of how new technologies can support mission-driven organizations and help them operate more efficiently. Watch the full video and read Project Open Hand’s story: https://lnkd.in/gN6Z5HM6 Big thanks to Boone Ashworth and WIRED for covering the story: https://lnkd.in/eGfzK8Vx) #robotics #physicalai #nonprofit

We’re excited to welcome Dmitriy Ganapolskiy to our manufacturing team as a Senior Technician! Dmitriy brings over a decade of engineering technician experience from companies like Tesla and Skydio, where he worked on scaling complex hardware systems. At Chef, he’ll play a key role in optimizing and scaling the production of our food robotics systems deployed across North America and Europe. We’re thrilled to have him on the team. Welcome, Dmitriy!

Ingredient onboarding is one of the hardest problems in physical AI. Food doesn’t behave like rigid objects; it’s messy, variable, and highly context-dependent. That makes traditional robotics approaches brittle and hard to scale. Our AI team has developed SAGE, an LLM-powered agent, to solve this. Instead of relying on simple similarity matching, SAGE combines: • Structured production data • Expert heuristics encoded in prompts • Real-time reasoning over ingredient behavior The result: a system that can recommend utensils and manipulation parameters for new ingredients faster, more consistently, and with full traceability. This is what physical AI looks like when it’s designed for real-world environments. Read more on our engineering blog: https://lnkd.in/ggTYEthV #physicalai #robotics #ai