Alex de Vigan’s last business got swamped by the rise of generative AI, but now he’s betting the tech that team had built, a platform for creating tempting product renderings for e-commerce, can be re-purposed to make robots more useful.
His new venture, Physicl, has raised $10 million in seed funding from Inovia Capital to help world model designers and robot builders solve thorny training problems with more realistic simulations.
Founder and CEO de Vigan spent the last decade building nfinite, which made it easy for brands and retailers to create 3D models of their goods and display them online. The arrival of transformer-based AI models led that company to focus more on generative imagery, but de Vigan saw a new purpose for his team and the tech stack they’d built to faithfully render 3D environments.
“What we were super unique about doing in 3D, AI was going to replace, but something was missing in AI: its ability to understand, interact, and manipulate space and objects,” de Vigan says.
Researchers developing physical AI models to make robots more capable of tasks like handling diverse, small objects have been frustrated by the returns of training them on video, and the expense of gathering data through remote operations.
Simulations in digital worlds are seen as a promising solution, but ensuring that these simulated worlds are faithful to real-world physics has proven difficult. de Vigan believes that the pipeline his previous company built to “transform unstructured, incomplete, heterogeneous real-world information into a high-fidelity physical representation… with correct geometry, materials, and physics” can now be used to make physical AI models more effective.
Physicl is entering a crowded category of startups vying to provide data and tooling to foundation model builders and robot operators. “The space is very young and and and very blurry sometimes,” de Vigan admits. There are a variety of companies collecting, labeling, or generating data in different modalities as their customers try to figure out the secret sauce that will make robots live up to their promise.
In particular, Physicl wants to focus on post-training models on difficult edge cases. The startup is releasing a product called Studio with 10,000 3D environments and 100,000 assets that represent various objects. It lets users create scenes to generate synthetic data for training robots on specific problems. For example, Studio could be used to generate training data to help a robot that has trouble discerning a cup on a cluttered counter-top, or while the cup is in shadow.
SpAItial, a world model builder, collaborated with Physicl as an early customer to improve the fidelity of its products, the latest of which, Echo-3, was released last week. SpAItial’s tool generates 3D simulations from 2D imagery, and to do that effectively, its CEO and founder Matthias Niessner said they needed to train on the kind of high-fidelity synthetic data Physicl offers.
However, Physicl wants to be more than a data provider; it sees itself as the tooling for ongoing synthetic data generation. “What we created 10 years ago, without knowing it, was basically an agentic workflow to transform unstructured information into high fidelity 3D,” de Vigan said. He imagines customers bringing trillions of images or millions of hours of LIDAR capture data to feed into his platform and turn them into high-quality 3D representations that can be used to generate billions of permutations to train their models.
“The real challenge is, you have to curate the data by hand, and this is a very long cycle,” Niessner, also a professor of robotics at the Technical University of Munich, told TechCrunch. “To us what seems very attractive is you can just do this on the fly […] You’re running Claude, and it checks, ‘Oh, I need more data of this in my training,’ and then it auto generates that data, depending on the training state. That, to me, is super exciting, and this is exactly what Physicl is offering.”
