The robot that learned to scoop: why food is the toughest test for physical AI
Chef Robotics has served 118 million portions in real kitchens and used that experience to build what it calls the world's largest dataset for robots handling soft, unpredictable materials. Now it wants to take those lessons far beyond food.
Key points
- Chef Robotics has completed over 118 million servings across more than a dozen food manufacturing facilities in North America and Europe.
- The company built its Food Foundation Model (FFM) on what it claims is the largest real-world dataset of robots handling soft, deformable materials.
- FFM lets robots generalise to new ingredients with minimal retraining, reducing the usual setup cost for each new food item.
- CEO Rajat Bhageria will present the company's findings at RoboBusiness 2026 on 20-21 October in Santa Clara, California.
- Chef Robotics says the techniques it developed for food could transfer to medicine, agriculture and flexible packaging.
Scooping a portion of chicken tikka masala sounds simple. For a robot, it is one of the hardest physical tasks in existence.
Every ingredient shifts, squishes and sticks differently. Weight changes batch to batch. Temperature affects how a sauce clings to a spoon. A robot that handles a firm carrot must instantly recalibrate for a soft slice of tofu. These are not edge cases. They happen thousands of times a day in any working kitchen or food factory.
Why is food so hard for robots?
Most robots train on rigid objects: boxes, car parts, metal components. Food breaks every assumption those training sessions build. Chef Robotics, a San Francisco company that automates food production lines, has been collecting data on exactly this problem since it began deploying machines in real factories.
The result is what the company calls the Food Foundation Model, or FFM. A foundation model is a large AI system trained on a broad dataset, similar in concept to the technology behind ChatGPT, but here built entirely around the physical task of handling food. The FFM was trained on data from 118 million real servings, which Chef Robotics describes as the largest dataset ever assembled for robots manipulating soft, deformable materials.
The practical payoff is speed of adaptation. Normally, teaching a robot to handle a new ingredient requires extensive reprogramming. With FFM, Chef Robotics says its robots can generalise to new items with far less retraining time, cutting the cost of adding a new product to a food line.
What does this mean beyond the kitchen?
The bigger claim is that food turns out to be a useful training ground for all kinds of physical AI challenges.
The techniques Chef Robotics developed, including what the company calls adaptive grasping (adjusting grip strength on the fly), tactile feedback integration (using touch sensors to sense pressure and texture), and training on huge variation in real conditions, apply wherever materials are soft and unpredictable. Think surgical instruments, flexible packaging, or fruit picking in agriculture.
CEO Rajat Bhageria will make that case at RoboBusiness 2026, a commercial robotics conference scheduled for 20-21 October in Santa Clara. His talk, first flagged by The Robot Report, is titled "Why Food is Physical AI's Hardest Problem and Most Promising Catalyst."
| Metric | Figure |
|---|---|
| Servings completed | 118 million |
| Facilities active | More than 12 |
| Regions covered | North America and Europe |
| Conference date | 20-21 October 2026 |
| Conference location | Santa Clara, California |
What should ordinary people take from this?
For most people, the direct effect is a more consistent ready-meal or hospital food tray. Robots trained at this scale make fewer scooping errors and maintain portion sizes better than a fatigued human worker on hour eight of a shift.
Longer term, the same technology that learns to handle a soft bread roll could one day assist in packaging a fragile medical device or harvesting a delicate crop. Food, of all things, may be the domain that teaches robots how to handle the physical world at large.
