Physicl raises $10M seed to turn 3D simulation into robot training data

A $10M bet on synthetic worlds for robots

Physicl, a new startup from serial founder Alex de Vigan, has raised $10 million in seed funding from Inovia Capital, TechCrunch reports. The company's pitch is that robots will only become genuinely useful if their models can be trained in simulations faithful enough to real-world physics — and that de Vigan's previous company spent a decade, perhaps accidentally, building exactly the technology needed to create them.

From e-commerce renders to physical AI

According to TechCrunch, de Vigan spent around ten years building nfinite, which helped brands and retailers produce 3D models of their products for online display. The rise of transformer-based generative models threatened to swamp that business, but de Vigan concluded that while AI could eventually replicate his team's rendering work, it still lacked something his technology handled well: an understanding of how to represent, interact with and manipulate physical space and objects.

The underlying problem is well known in robotics research. As TechCrunch explains, teams developing physical AI models for tasks such as handling small, diverse objects have been frustrated by the poor returns from training on video, while gathering data through remote teleoperation is expensive. Simulated worlds look like the answer, but only if their geometry, materials and physics match reality closely enough that lessons learned inside them transfer outside.

De Vigan believes nfinite's pipeline — which converted unstructured, incomplete real-world information into high-fidelity physical representations — can be repurposed for exactly that job.

Studio launches with 10,000 environments

Physicl's first product is Studio, shipping with 10,000 3D environments and 100,000 object assets. Users assemble scenes to generate synthetic data aimed at a robot's specific weaknesses. The company's examples involve a model that struggles to recognise a cup on a cluttered countertop, or the same cup sitting in shadow — edge cases that are hard to film in the real world but straightforward to render in a controllable simulation.

That focus on difficult edge cases during post-training is where Physicl wants to concentrate, rather than competing directly on foundation-scale pre-training.

The startup is candid about the competition. De Vigan admits to TechCrunch that the category is young and often blurry, with a crowd of companies collecting, labelling or generating data in different modalities while customers search for the formula that makes robots deliver on their promise. His differentiation is positioning Physicl not as a dataset vendor but as tooling for continuous synthetic data generation. He describes what nfinite built as, in effect, an agentic workflow for turning unstructured information into high-fidelity 3D, and he imagines customers bringing trillions of images or millions of hours of LIDAR capture to the platform and receiving billions of training permutations in return.

SpAItial signed on as an early customer

World model builder SpAItial worked with Physicl as an early customer to improve the fidelity of its products, including Echo-3, released the week before the funding announcement. SpAItial generates 3D simulations from 2D imagery, and its founder and CEO Matthias Niessner — also a robotics professor at the Technical University of Munich — told TechCrunch that doing this well required training on the kind of high-fidelity synthetic data Physicl provides.

Niessner's argument is about removing humans from the data loop. Hand-curating training data is a slow cycle, he says, whereas the attractive scenario is on-the-fly generation: an agent such as Claude, running alongside a training job, notices it needs more examples of something, and the platform automatically produces them based on the current training state.

Why it matters

The consensus bottleneck in embodied AI is increasingly data, not model architecture — real-world demonstrations are slow and costly to collect, and video alone hasn't been enough. Physicl's seed round is a concrete bet that synthetic environments can close that gap, and that a decade of e-commerce 3D tooling transfers cleanly to robotics. If Niessner's vision of self-directed data generation pans out, training data becomes something models request and receive in real time rather than something humans assemble by hand. That said, the market is crowded and unproven, and the field's own founders describe it as blurry — so the $10 million buys Physicl a chance to prove fidelity, not a settled category.