Chef Robotics

How many auger fillers does it take to fill a three-ingredient tray? Three, as each filler is calibrated for one ingredient. High-mix meal assembly requires equipment to handle many ingredients, constant SKU changes, quick changeovers, and zero tolerance for giveaway. 𝗛𝗼𝘄 𝗮𝘂𝗴𝗲𝗿 𝗳𝗶𝗹𝗹𝗲𝗿𝘀 𝗯𝗲𝗵𝗮𝘃𝗲 𝗼𝗻 𝗮 𝗵𝗶𝗴𝗵-𝗺𝗶𝘅 𝗹𝗶𝗻𝗲 → Each filler has custom hardware for one ingredient; running dozens of SKUs means dozens of single-purpose fillers sitting idle between runs → Changeovers take hours, cleaning and recalibrating between every SKU; Allergen switches take even longer, since fine powder residue can become airborne 𝗪𝗵𝗮𝘁 𝗮𝘂𝗴𝗲𝗿 𝗳𝗶𝗹𝗹𝗲𝗿𝘀 𝗮𝗿𝗲 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗯𝘂𝗶𝗹𝘁 𝗳𝗼𝗿 → Built to run one dry, free-flowing ingredient, like spices, flour, or protein powder, at high volume, reliably 𝗪𝗵𝘆 𝘄𝗲 𝗯𝘂𝗶𝗹𝘁 𝗖𝗵𝗲𝗳 𝗿𝗼𝗯𝗼𝘁𝘀 → Chef's AI-enabled meal assembly robots pick and place ingredients by weight, not volume, so weight never drifts the way a rotation count would → Changeovers take minutes, not hours, because there's no hopper to clear or screw to recalibrate → One robot handles every ingredient in a SKU, wet or dry, that would otherwise need a separate filler, or a person, for each one Full comparison, including where an auger filler is still the right call: https://lnkd.in/gz8aAYj6 #Foodmanufacturing #Aiandrobotics #Foodautomation #augerfillers

We’re excited to welcome Ron van Thiel to Chef! Ron is joining us as a Staff Mechanical Engineer. With a mechanical engineering degree from the University of California, Berkeley, and decades of work experience, most recently at Bear Robotics and EverCharge, Ron brings deep robotics and food industry experience, helping us design faster, more reliable robots for the food industry. If this sounds interesting to you, see our open roles: https://lnkd.in/g2J6g63b

What if robots could imagine what happens next before they act? That's the promise of world models: AI systems that learn to predict possible futures, plan, learn, and generate new experience before taking an action in the physical world. In our latest survey paper, we map the rapidly evolving field of world models and world-action models. We cover over 220 research contributions and look at 160 systems across six modeling approaches and six application domains, from robotics and autonomous driving to reinforcement learning, gaming, and general-purpose simulation. One of the biggest open questions is whether models that generate a convincing future actually understand the physical world. That distinction is critical for robotics: a future that looks plausible but violates physics isn't good enough when a robot needs to act on it. Read our paper for an accessible overview of the field, the major approaches, and the research challenges that need to be solved before we can bring world models into the physical world: https://lnkd.in/g8m46i-7 #robotics #worldmodels #physicalai

Given how fast AI research is evolving, it might be surprising how relevant papers published a decade ago can still be today. Congratulations to our Senior Staff AI Research Scientist, Inkyu Sa, whose 2016 paper, DeepFruits, was recently recognized by Sensors MDPI as one of the journal's Most Influential Papers (2001–2026): https://lnkd.in/eVfEHsxa Inkyu's paper focused on helping agricultural robots detect fruit using deep neural networks. While this research was applied to autonomous harvesting, the underlying challenge is one we're still working on today: enabling robots to perceive and interact with highly variable food items in the real world. At Chef, those foundational ideas have evolved into our work on the Food Foundation Model and our bi-manual physical AI system that can assemble meals such as burgers and burrito bowls. Instead of recognizing fruit on a plant, we're teaching robots to recognize ingredients in tubs and pans, manipulate them precisely and reliably, and generalize learnings to new recipes and use cases. Compared to 10 years ago, AI models are more capable today, datasets are larger, and applications have expanded. But the fundamental challenges of perception and food manipulation remain surprisingly similar. Congratulations, Inkyu, on this well-deserved recognition! Read his paper at https://lnkd.in/gw94Wv3

Most companies raise capital to find product-market fit. At Chef Robotics, we did it the other way around. From day one, our CEO, Rajat Bhageria, insisted on this: don't go to investors with a vision alone. Bring evidence from customers. Signed commitments. Willingness to pay. Proof that the problem is real, not just interesting. That sequencing runs against the standard playbook, where most companies raise first and look for proof after. We wanted proof first, so that when we did raise, investors weren't asked to believe in a hypothesis. Proving traction first is slower in the short term in a category as capital-intensive and unforgiving as robotics. But eventually it compounds: Chef robots have now completed over 127 million servings in production, with customers seeing 2-3x output increases and changeovers under 10 minutes. More of this in the TechCrunch Disrupt conversation: https://lnkd.in/gxhfzj8d #techcrunchdisrupt #aiandrobotics #foodautomation #physicalAI

What does it take to build a lasting food business? In Episode 2 of our Food Builders podcast, Rajat sits down with Sameer Malhotra, Co-Founder and CEO of Cafe Spice, to discuss the company's journey from a family-run restaurant business in New York City to a leading producer of ready-to-eat meals. Sameer shares how he and his father, Sushil, scaled Cafe Spice by expanding into new markets, launching new product lines—including Latin American meals under the Cantina Latina brand—and building a strong team along the way. They also discuss leadership, hiring, company culture, and what the future holds for Cafe Spice and the food industry. 🎙️ Watch or listen to the full episode on YouTube: https://lnkd.in/gfra4caX or Spotify: https://lnkd.in/g6wwmTNH

We just published a new survey paper on vision-language-action models (VLAs) for bimanual manipulation. VLAs have gained popularity in robotics research, but while they've demonstrated impressive capabilities, it can be difficult to understand how different architectures compare and which ones are best suited for real-world deployment, especially for bimanual manipulation. To answer these questions, we reviewed more than 200 papers and compared 31 VLA methods. We examined their architectures, training strategies, action representations, and real-world applications. Our key takeaway is that not all bimanual tasks require the same level of coordination, which impacts the best VLA method and architecture to use. For tasks that require two robot arms to remain tightly synchronized, architectures that generate both arms' actions jointly are better suited than approaches that generate them sequentially. Among today's methods, flow-based action generation offers one of the strongest reported combinations of coordinated action generation and the speed required for real-time control. The survey also examines where the field needs to go next, including: 1. Standardized benchmarks for bimanual manipulation 2. Better force, tactile, and multimodal sensing 3. The safety and reliability needed for large-scale commercial deployment Read the full survey paper: https://lnkd.in/gWYmtwQH #robotics #physicalai

A food depositor deposits one ingredient. A Chef robot assembles an entire meal. Here's when each one is the right fit. Depositors (also called dispensers or fillers) are excellent at depositing one flowable ingredient like sauces or dressings at high speed. But they fall short for high-mix meal assembly for two reasons: → A depositor is unable to handle many ingredients such as diced proteins, vegetables, grains, and mixed foods, as they don't flow well through a nozzle. → Even for ingredients a depositor can handle, like sauces, changeover time makes it impractical. Switching ingredients means cleaning the full ingredient contact path, which takes hours when SKUs switch several times a day. We built Chef robots to solve this. Here's a buying guide comparing the two in detail: https://lnkd.in/gdebFNng #foodmanufacturing #foodautomation #mealassembly #foodrobotics

How should robots learn when things don't go as planned? Our latest engineering blog introduces ValueFormer, a lightweight critic model that helps our Food Foundation Model learn from mistakes, not just successes. ValueFormer watches the same camera feeds as our robot, evaluates progress in real time, and provides feedback that behavior cloning alone can't. By incorporating human interventions and critic-generated feedback into training, we improved successful burger assembly from 70% to 85% on a challenging consecutive assembly task. This work builds on our broader vision for physical AI: combining production data, human expertise, and continuous learning to create robots that improve over time. Read the full tech blog: https://lnkd.in/gNd4Zs2v

Every meal makes the next one better. Chef robots have assembled more than 120 million servings in production, generating one of the world's largest datasets for food manipulation along the way. Every serving adds production data that improves our physical AI models—helping Chef robots handle more ingredients, more tray types, and more edge cases, all on the same hardware. This data flywheel has been turning since our first deployment. Today, it helps customers: - Reduce giveaway by up to 88% - Increase output by 2–3× - Improve labor productivity by up to 60% - Reach full production faster than earlier deployments Unlike language AI, there is no internet-scale dataset for food manipulation. The only way to build these models is through years of production experience on real manufacturing lines. In our latest blog, we explain how this flywheel works, why it compounds over time, and why we believe production data is one of the strongest moats in physical AI. Read more: https://lnkd.in/ggRnZWbc