Nomagic’s new AI lab headed by former Google DeepMind researcher claims success in early deployment of ‘AI brain’ for warehouse robots

Nomagic says the company's new 'vision-language-action' model cut robot errors requiring human intervention in half at logistics customers.

“Embodied AI” and “physical intelligence” are all the rage with Silicon Valley investors these days. The idea is that AI’s next frontier will be systems that don’t just use software but can take action in the real world through robotic devices, from self-driving cars to humanoid robots.

Many startups are chasing AI models that can serve as general-purpose “robot brains,” able to be dropped into any kind of robot and told to do almost anything. This is a shift from the kinds of systems that traditionally controlled industrial and warehouse robots. This control software often required weeks or months of on-site programming to perform even one task well.

Still, most of these general purpose AI models perform significantly below human-level accuracy on each task, at least right out of the box. The hope is that with just a little bit of additional on-site, task-specific training, these robots will eventually be able to master that task—reducing the barriers to deploying robots in many sectors.

Nomagic, a company with European headquarters in Warsaw, Poland, and U.S. headquarters in Sandy Springs, Georgia, is pursuing a different approach: rather than going from generality to task-specific mastery, it is creating AI robot brains that are extremely accurate at specific tasks right out of the box, and then hoping to eventually build from mastery of these individual tasks towards a general purpose system.

To pursue this goal, earlier this year Nomagic created an AI research lab led by Markus Wulfmeier, a former Google DeepMind robotics researcher, who now serves as Nomagic’s chief scientist. Now Nomagic has announced that it has deployed its first vision-language-action (VLA) model—a type of AI model that can perceive objects in the world, receive and understand text-based instructions from people, and then take actions in the world—to paying customers. The company says it is among the first companies in the world to run VLAs in a live production environment, rather than in lab experiments or staged demos. 

The early results, according to the company, are tangible if unglamorous: by aiming the VLA at the most common “edge cases” for its warehouse robots—somewhat uncommon situations where a robot gets stuck and has to call for human assistance—Nomagic says it has roughly halved the rate of these robot-caused interventions in live operations.

Nomagic’s first VLA deployment is with Brack.Alltron, Switzerland’s second-largest e-commerce platform, which has been using robots from Nomagic to automate order picking and packing in its warehouses. Roland Brack, the company’s founder and owner, said the addition of Nomagic’s VLA systems marked a step change.

“In the past, our goal was simply to minimize manual intervention. Today, we are seeing robots that truly understand their environment,” he said. “This intelligence allows us to run autonomous shifts through nights and Sundays, ensuring we stay ahead of peak demand without increasing the pressure on our human workforce.”

Nomagic’s concedes though that its VLA system is not perfect, even at the specific box picking tasks the company is targeting. “Our VLAs aren’t at 99.9% success on their own yet — no one’s customer-deployed VLAs are there yet,” the company said. But it says it has created a system around the VLA: Nomagic’s older “classical” robotics software acts as a “harness,” catching errors and enforcing safety, so the entire system can be trusted in a customer’s warehouses.

“The bar in the physical world is high: 99.9% [reliability] isn’t a marketing number, it’s the cost of being allowed in the building,” Kacper Nowicki, Nomagic’s cofounder and CEO, said. “So we built a harness that clears it from day one, while the AI inside keeps getting better.” Over time, both Nowicki and Wulfmeier said they expect stronger models to gradually make parts of that harness unnecessary, just as has begun to happen with digital AI.

Nomagic recently won the 2026 International Intralogistics and Forklift Truck of the Year (IFOY) Award for Shoebox Picker, which goes to the company whose sorting and picking device can master a notoriously difficult challenge in warehouse automation: handling two-piece shoeboxes without the lids falling off.

A former core member of the Gemini Robotics team at DeepMind, Wulfmeier frames Nomagic’s approach as a deliberate contrast with the prevailing strategy of competing embodied AI labs.

“Most of our community is racing to build the most general robot brain,” he told Fortune. “We’re betting that the harder part is actual mastery and that it has to be earned in real deployments first.”

Wulfmeier said that the physical world is dominated by a very long tail of rare situations. This is the same problem that has caused the roll-out of autonomous vehicles to be much slower than many anticipated a decade ago. The AI models running those vehicles have to be trained for an endless array of edge cases.

Today, most companies working on the AI models for robotics train either in simulation and then transfer those skills to real world settings (which roboticists call “sim-to-real” training), or by having humans initially operate the robots by remote control, creating examples that the robot learns to imitate. Some combination of those two methods can get an AI model to 80% performance accuracy on a fairly wide array of tasks, Wulfmeier said. But, working in a real warehouse, 80% is basically useless, he said. If a robot needs a human to step in even once an hour, the economics of automation often collapse.

Wulfmeier did extensive “sim-to-real” work at DeepMind and said he still believes in simulation and uses it in parts of Nomagic’s own pipeline. But he said he doubts either simulation or human teleoperation can economically close the remaining gap to the level of reliability the physical world demands.

Nomagic said that one major advantage it has over pure research labs is that it is able to gather tons of real world data from the fleet of robots the company already has deployed with customers. That existing fleet generates millions of successful package picks every month (two million with the fashion platform Zalando alone, the company says), and that stream grows as more robots are deployed. Rather than relying primarily on teleoperation or simulated environments, Nomagic trains its VLAs on this deployment data, which Wulfmeier describes as unusually rich and diverse.

Tristan d’Orgeval, Nomagic’s co-founder and chief strategy officer, said deploying robots to the real world first is a key differentiator between Nomagic and competing companies building AI systems for robots. “We didn’t build a lab and then go hunting for a problem,” he said. “We started in real operations, with customers who need our robot, and capable AI emerges out of that. The order matters — it’s what separates a demo from a business.” 

This story was originally featured on Fortune.com