Nomagic Names DeepMind's Wulfmeier Chief Scientist for VLA Models

Warsaw-based warehouse robotics firm Nomagic has poached Markus Wulfmeier from Google DeepMind as its first Chief Scientist, tasking him with building the company's Robotics Foundation Model from scratch. The hire was reported Thursday by Business Insider Markets, framing it as a defining step in the company's push to scale physical artificial intelligence across warehouse fulfillment.

Wulfmeier arrives at a moment when the research community is actively debating whether synthetic simulation or real operational data will define the next generation of robot learning. Nomagic has already placed its bet. The company has spent years building what it calls the "Library of Chaos," a proprietary dataset of millions of edge cases harvested continuously from live warehouse deployments.

The data bet

The analogy Wulfmeier reached for in his public statement is telling. He described Nomagic's production data pipeline as the "internet data of robotics," echoing how large language models became broadly capable by training on the sprawling, unfiltered breadth of human-generated text rather than curated corpora. Robots already generating commercial value, he argued, simultaneously generate training signal for the models that will succeed them.

This contrasts sharply with approaches built on teleoperation demonstrations or synthetic environments. Both can bootstrap capabilities quickly in controlled settings, but they tend to fail at the distribution tail: cluttered scenes, unusual object geometries, and cascading failure modes that live fulfillment floors surface daily. According to the announcement, Wulfmeier will lead end-to-end vision-language-action (VLA) model development, improve complex object manipulation pipelines, and extend both online and offline reinforcement learning across Nomagic's software stack.

Why this hire matters

VLA models have become one of the most competitive research fronts in physical artificial intelligence. The architecture class conditions robot actions jointly on visual observations and natural-language instructions, with substantial work coming from DeepMind, Physical Intelligence, and leading academic groups in recent years. Recruiting directly from DeepMind's robotics research brings institutional expertise that few industrial robotics firms can access.

What Nomagic offers in return is something most research labs cannot match: continuous deployment scale. Production environments generate training data not in the thousands of curated demonstrations typical of academic settings, but in the millions of real interactions that include the messy, ambiguous scenarios where learned policies tend to break. CEO Kacper Nowicki, quoted by Business Insider Markets, framed VLA models as undergoing the same transformation large language models completed over the past decade.

That comparison deserves scrutiny. Generalization in the physical world is considerably harder to measure than performance on text benchmarks, and success in warehouse manipulation does not automatically imply broader cross-domain robotic capability. Still, the core claim aligns with what the broader artificial intelligence literature has documented across language and vision domains: data diversity and operational scale drive capability gains that narrow benchmarks routinely underestimate. Nomagic is not alone here; Physical Intelligence, Covariant, and Apptronik are all pursuing foundation model strategies, each with a different view on how heavily to lean on simulation.

What to watch

No architecture details, benchmark results, or deployment timelines accompanied the announcement. Wulfmeier's stated intent to engage with the wider research community suggests some findings may eventually surface in preprints or conference submissions. Those disclosures will be the first real test of whether the production-data thesis produces measurable gains over simulation-trained baselines.

The deeper question for the field is whether warehouse-scale deployment displaces academic benchmarks as the primary proving ground for physical AI. If Nomagic's approach works, the next milestone in robot generalization may not appear in a paper at all, but in a fulfillment center's throughput numbers.

Frequently asked questions

What is a Vision-Language-Action (VLA) model in robotics?

A VLA model jointly conditions robot action generation on visual observations and natural-language instructions within a single learned architecture. Unlike earlier modular designs that separate perception and planning, VLA models aim to generalize across novel tasks without task-specific engineering.

What is Nomagic's "Library of Chaos"?

It is Nomagic's proprietary training dataset, built from millions of real-world edge cases collected continuously during live warehouse operations. The name reflects a focus on capturing the unpredictable, cluttered conditions that researcher-designed scenarios routinely miss.

How does Nomagic's approach differ from simulation-based robot training?

Simulation is cheap but struggles to close the gap with real-world conditions at the tail of the distribution. Nomagic relies on production robots as a continuous data source, aiming to reduce the simulation-to-real transfer failure that limits many deployed systems.

Which companies compete in the robotics foundation model space?

Physical Intelligence, Covariant, and Apptronik are among the prominent players pursuing foundation model strategies for robotic manipulation. Each takes a different stance on the simulation-versus-real-data tradeoff, and no clear winner has emerged at production scale.