Fizyr’s Vision AI automates “edge cases” in logistics
At NCAS’26, Ken Fleming demonstrated how artificial intelligence finally gives robots the crucial ability to “see” and reason.
Fizyr’s development began as an incubator project at Delft University of Technology, where the team won the Amazon Picking Challenge by both recognizing and placing objects. However, an important business lesson soon followed: building and maintaining physical robot fleets is fundamentally different from developing algorithms. To scale up, the company chose to focus entirely on what it was truly good at—computer vision software—and leave the hardware side to integration partners.
Today, Fizyr provides the “thinking part” that translates raw camera data into concrete instructions for robots that are, in and of themselves, “dumb.” At the National Congress on Autonomous Systems (NCAS’26), Fleming demonstrated how AI gives robots the essential ability to see and reason.
Gaining Control Over Chaos and Quality Control
While many automated warehouses rely on robots that move neatly stacked boxes, Fizyr focuses specifically on the messy “edge cases.” Think of chaotic piles of packages that have been stepped on, or products lying in random orientations. It is precisely in these situations—where traditional programming fails—that Fizyr’s Vision AI excels.
Fleming also emphasized a second, often underestimated function of Vision AI: quality control. The system not only determines where a robot should pick up a product but also assesses its quality. In food processing, for example, this means that the AI not only locates a bell pepper but also analyzes any damage, assesses its severity, and compares it to the acceptance criteria of specific retailers before deciding whether the product can be packaged.
Heerenveen as a breakthrough: sorting the unsortable
A striking example in the presentation came from the collaboration with Pallet Sorting Systems in Heerenveen. Sorting wooden pallets is notorious: it’s heavy work, involves splinters, and is extremely monotonous. Many employees quit within six months, leading to a constant cycle of recruitment and training.
Automation seemed like a logical step, but turned out to be extremely complex. The global supply chain uses a huge variety of pallets: different sizes, varying dimensions, weights, and moisture content.
Initially, Fizyr’s engineers tried to train the AI based on strict industry standards with a 2% tolerance. That failed completely. The system couldn’t handle the enormous variation found in real-world situations—something that human workers do intuitively.
The missing link: “Arie,” the human expert
The breakthrough came when Fizyr realized they lacked domain expertise. They brought in an experienced pallet sorter: Arie. He had an intuitive understanding of what constitutes a good or bad pallet, built up through years of hands-on experience.
By having Arie work alongside the software engineers and provide his input on the AI’s decisions, the team was able to translate human intuition into the neural network.
The result: a fully automated sorting tunnel with seven cameras. Pallets are fed into the system; the cameras send data to Fizyr’s AI brain; and the system immediately determines what the object is and where it needs to go.
Hardware-independent and ready for the next step
To ensure maximum flexibility, Fizyr deliberately remains independent of specific cameras and robots. This allows integration partners to choose the best hardware for their application themselves.
In addition, the company is collaborating with NVIDIA, utilizing Jetson platforms. This reduces the heavy computational load on large GPU servers, which is essential for battery-powered, mobile robots.
The collaboration in Heerenveen is now being expanded further. While the AI currently sorts by type and detects heat treatment, the system will also perform full inspections early next year, including the automatic detection of broken boards and protruding nails.
For Fleming, the most important lesson for the Autonomy Economy is clear: true innovation only happens when the best software engineers, the best machine builders, and the end users come together in a single integrated solution.