We raised $10M led by Inovia Capital, with US Venture Partners, and today we launch Studio, our first app: describe the situation your model gets wrong, and generate it thousands of ways with exact ground truth on every frame.
Every leap in AI has followed a leap in data. Language models learned from the text of the internet, image models from billions of captioned photos. Physical AI, the robots, autonomous systems and world models that have to act in real space, has no such corpus. The physical world does not come as a dataset.
The gap shows up at the end. A robot that succeeds nine times out of ten makes a great demo, but when the technology operates alongside people in a kitchen, warehouse or hospital, no one can deploy a model that fails every tenth run. These last points of reliability come from rare situations: the cup half hidden by a bag, the shadow that crosses the counter at sunset, the box stacked slightly off. Broad deployment hinges on physical AI’s ability to act reliably within these edge cases, but they are almost impossible to capture in the real world, and when you do catch one, you cannot replay it or vary it.
I have spent the last ten years digitizing the world in 3D. With my team, we built a library of real objects and rooms, and the software to vary them, for some of the world's largest retailers.
Then a different kind of team started calling: robotics labs, world model labs, vision teams. They did not need prettier pictures. They needed the exact situations their models failed in, with every pixel labeled, and they needed thousands of them.
That is when it clicked: physical worlds cannot be scraped. They have to be built, from geometry and physics that are known, so every label is exact and every condition can be changed on purpose. Physicl exists to give every physical AI team the worlds its models need to learn from, starting with the cases they get wrong.
Today, we are announcing a $10M seed round led by Inovia Capital, with participation from US Venture Partners, to address physical AI’s edge-case problem. It funds the team in Paris and the San Francisco Bay Area, and the platform behind Studio.
"We believe physical AI represents one of the most significant emerging opportunities in technology, but the industry still faces fundamental challenges in translating model capabilities into real-world performance. Physicl is tackling a critical piece of that puzzle: scalable, high-quality physical data grounded in real-world physics."
Steve Woods, Partner at Inovia Capital
We work with advisors who built the field we are building for: Thomas Wolf (Hugging Face), Christoph Lassner (World Labs), Matthias Niessner (TUM), Yaser Sheikh and Sébastien Deguy.
Studio lets researchers recreate the situations their models get wrong, generate as many variations of them as they need, and train on them. Three things set it apart.
That is what the long tail needs: not more average data, but the rare case on demand, labeled exactly, as many times as training needs. Inside Studio covers how it works in detail.
SpAItial, which builds world models that turn real rooms into explorable 3D, leveraged Physicl data capabilities and more than halved its reconstruction error on real rooms.
"Physicl gives any researcher access to virtually any physical training data they can imagine. That changes what a small team can build in physical AI, and we saw it in our results right away."
Matthias Niessner, CEO of SpAItial and Professor of Computer Science at TUM
Every model that works in the real world got there the same way: it saw enough of the moments that matter. In physical AI, the moments that matter most are often the rare ones, the mug behind the bag at sunset, and nobody has been able to supply enough of them.
That is what we are building Physicl for: the worlds physical AI learns from, so any research team can make its models dramatically better, starting with the cases that break them today.
Studio is the first step, and there is a lot more to come.
Studio is in beta. We're opening it to research teams a few at a time, free during the beta. Bring the situation your model gets wrong, and generate the world that teaches it.
