The 3D rendering company e-commerce didn't need anymore just raised $10M to teach robots the real world
Alex de Vigan's last startup made 3D product shots for online stores until generative AI ate that business. His new one, Physicl, is selling the same 3D pipeline to robot builders as training data, backed by a $10 million seed round.
Alex de Vigan spent a decade building nfinite, a company that made 3D renderings of products so online stores didn’t need a real photo studio. Then generative AI models got good enough to fake a product photo from a text prompt, and that business got, in his own word, swamped. So on November 6 he announced a new startup that points the same pipeline somewhere else: training data for robots.
De Vigan’s new startup, Physicl, announced a $10 million seed round led by Inovia Capital on November 6. The pitch isn’t a new idea, exactly. It’s an old 3D-rendering problem wearing a robotics company’s clothes.
What Physicl is actually selling
Robots that have to grab a cup off a cluttered counter, or work out which drawer is which, learn from two kinds of data: real-world video, which is slow and expensive to collect, or simulated 3D worlds, which are fast but only useful if the physics and the geometry are right. De Vigan’s bet is that the pipeline nfinite built to turn messy real-world scans into accurate 3D product models can do the same job for robot training scenes.
Physicl’s first product, called Studio, ships with 10,000 pre-built 3D environments and 100,000 object assets. A team describes the scene it needs, and Studio generates images and video from it, with depth maps, segmentation and camera pose attached to every frame automatically. Change the lighting or move one object, and the rest of the scene stays put, which is the part you can’t do with a single real video clip.
| Data source | Cost to collect | Labels | Repeatable |
|---|---|---|---|
| Real-world robot footage | High (remote ops, hardware time) | Hand-labeled, often approximate | No |
| Physicl Studio scenes | Compute only, after the $10M build-out | Depth, segmentation, pose on every frame | Yes, change one variable at a time |
My read: the framing that matters here isn’t “robots,” it’s “labels you don’t have to make by hand.” Ground-truth labeling is the expensive, boring part of training any model that has to perceive a 3D world, and that’s the problem Physicl is actually pricing itself against.
A real customer, a smaller claim than it sounds
SpAItial, a world-model startup led by Matthias Niessner, a robotics professor at the Technical University of Munich, is an early Physicl customer. SpAItial’s own product, Echo-3, turns 2D images into 3D simulations, and Niessner told TechCrunch his team needed higher-fidelity synthetic data to train it well. “The real challenge is you have to curate the data by hand, and this is a very long cycle,” he said. He described wanting a system where a model in training can flag what kind of data it’s missing and have Physicl generate more of exactly that, on the fly.
That’s one named customer with one on-record quote, not an independent benchmark of Physicl’s output quality. It’s a real endorsement from someone with a reason to know, not proof the product works at scale.
A crowded, blurry category
De Vigan isn’t pretending Physicl has the field to itself. “The space is very young and very blurry sometimes,” he told TechCrunch, pointing at the range of startups collecting, labeling or generating training data for robot and world-model builders in different ways. Physicl’s bet is narrower than “we do synthetic data”: it wants to be the ongoing tooling layer that keeps generating new edge cases as a model’s training reveals gaps, not a one-time dataset drop.
My take: a $10 million seed round is a modest bet in a field this crowded, and the company’s own advisors list (Hugging Face’s Thomas Wolf, World Labs’ Christoph Lassner, Meta Reality Labs’ Yaser Sheikh) reads more like a signal that the right people think the problem is real than proof Physicl has already solved it. The round only covers the next stage of proving the pipeline works outside de Vigan’s own old business.
Related: another AI-adjacent funding story broke the same week, when Palo Alto Networks reportedly offered $500 million for an AI agent security startup that neither side has confirmed. On the hardware side of the same problem, see Ring’s new security drone and patrol robot, which Amazon says learns a floor plan the same way Physicl’s customers train on synthetic scenes: by mapping a real space first.
Sources
Note on sources: facts about the funding round, Studio’s feature set and the Niessner and de Vigan quotes come from TechCrunch’s reporting, which includes on-record interviews with both. Physicl’s own site confirms the product description, founding date and advisor list, but does not mention the funding round itself.
- Physicl raises $10M to make the digital world real enough for robots - TechCrunch, November 6, 2026
- About Physicl - Physicl’s own company page