Nomagic, a Warsaw-based warehouse robotics firm, put a vision-language-action model into live customer operations. It says the model roughly halved the rate at which its robots stall for human help. Its new AI lab, run by a former Google DeepMind researcher, is betting on mastery before generality.
A Polish robotics company has quietly done the thing most robot labs are still demoing. Nomagic says it has deployed a vision-language-action (VLA) model into live warehouses with paying customers, Fortune reports. It says the move cut robot-caused human interventions by about half. The firm claims it is among the first to run VLAs in real production, not a staged demo.
Nomagic keeps its European headquarters in Warsaw and a US base in Sandy Springs, Georgia. Earlier this year it started an AI research lab. It hired Markus Wulfmeier as chief scientist. He is a former Google DeepMind researcher and a core member of the Gemini Robotics team.
A VLA is a single model that can see objects, read plain-language instructions, and then act. Many labs now chase it for embodied AI. The idea is that software should move things in the real world, not just answer questions on a screen.
Mastery before generality
Here Nomagic breaks from the pack. Most of the field races to build one general robot brain that drops into any machine. Nomagic runs the logic backwards. It aims for models that nail one task out of the box, then builds towards a general system.
“Most of our community is racing to build the most general robot brain,” Wulfmeier told Fortune. “We’re betting that the harder part is actual mastery and that it has to be earned in real deployments first.”
His reasoning rests on the long tail. The physical world throws up endless rare situations. That is the same problem that slowed self-driving cars. Training in simulation or by remote control gets a model to roughly 80 per cent accuracy. In a warehouse, 80 per cent is useless. If a robot needs a human once an hour, the economics collapse.
The harness
Nomagic admits its VLA is not perfect. “Our VLAs aren’t at 99.9% success on their own yet,” the company said, adding that no rival’s deployed VLAs are either. So it wraps the model in older “classical” software. That layer catches errors and enforces safety.
“The bar in the physical world is high: 99.9% isn’t a marketing number, it’s the cost of being allowed in the building,” said co-founder and chief executive Kacper Nowicki. “So we built a harness that clears it from day one, while the AI inside keeps getting better.”
The firm says its edge is data. Its deployed fleet already makes millions of successful picks each month. Two million come from the fashion platform Zalando alone. Nomagic trains its VLAs on that live stream, not on simulations.
Why it matters
The first VLA deployment runs at Brack.Alltron, Switzerland’s second-largest e-commerce platform. Founder Roland Brack said the shift was real. “Today, we are seeing robots that truly understand their environment,” he said. “This intelligence allows us to run autonomous shifts through nights and Sundays.”
The claims are Nomagic’s own. VLAs still fall short of full reliability on their own. Yet the deploy-first stance marks a real contrast with labs chasing general-purpose humanoids. It also raises the stakes in the race to automate Europe’s warehouses, a busy front in European robotics. As co-founder Tristan d’Orgeval put it: “We didn’t build a lab and then go hunting for a problem. The order matters. It’s what separates a demo from a business.”
