Physicl raises $10m to generate the rare scenes that trip up robots

Physicl raises $10m to generate the rare scenes that trip up robots

9 November 2026: Physicl says it has raised $10 million in a seed round led by Inovia Capital, with US Venture Partners taking part, to build synthetic training data for robots, autonomous systems and world models. The money pays for its team in Paris and the San Francisco Bay Area and for the platform behind Studio, an app it has opened in beta, free of charge, to research teams in waves.

Not yet independently verified. The amount and the investors rest on Physicl's own post of 6 November 2026, and we have not read a post from either investor. The SpAItial result is also stated by Physicl, citing an evaluation by a team led by SpAItial's own chief executive. We will update this when it can be confirmed, and remove this note.

The ninth success does not ship a robot

Alex de Vigan, the founder and chief executive, argues that physical AI lacks what language models had in the text of the internet: a ready-made body of data. "A robot that succeeds nine times out of ten makes a great demo," he writes, but nobody can deploy one that fails every tenth run alongside people in a kitchen, warehouse or hospital.

Those last points of reliability, in his telling, depend on rare situations that are hard to capture and impossible to replay: a cup half hidden by a bag, a shadow crossing a counter at sunset, a box stacked slightly off. Physicl's answer is to build such scenes in 3D from known geometry and physics, so that each can be repeated and changed on purpose.

Labels that come from the scene itself

A team using Studio describes the situation its model gets wrong and generates it thousands of ways. Depth, surface normals, segmentation, bounding boxes and camera pose are taken from the scene and not estimated from the rendered picture, which the company says keeps them accurate in dark corners and behind clutter, where models tend to fail. Light, clutter, camera position and objects can be changed one at a time or together.

Physicl says Studio starts with 10,000 environments and 100,000 underlying assets built from real geometry with known materials. Its intended users are labs, universities and companies building vision, world model and robotics systems.

The one customer result it gives comes from SpAItial, which turns real rooms into explorable 3D worlds. Physicl says SpAItial more than halved its reconstruction error on real rooms after training on its data. Matthias Niessner, who is both chief executive of SpAItial and a professor of computer science at TUM, also advises Physicl, and the evaluation was carried out by his team at TUM.

From retail 3D catalogues to robotics labs

Mr de Vigan says he spent ten years digitising the world in 3D, including a library of real objects and rooms for large retailers, before robotics and world model labs began asking for the exact situations in which their models failed. The company's site puts the team at about 30 people, with offices in New York, San Francisco and Paris, and says it launched publicly in March 2026 at NVIDIA GTC.

"Physicl is tackling a critical piece of that puzzle: scalable, high-quality physical data grounded in real-world physics," said Steve Woods, a partner at Inovia Capital. Other advisers named are Thomas Wolf of Hugging Face, Christoph Lassner of World Labs, Yaser Sheikh and Sébastien Deguy.

The post gives no valuation and no pricing for the period after the beta. Our robot manufacturer category lists 62 companies, and Physicl now has a profile of its own as a supplier to that field.

Company profile on TrustList: Physicl

Related on TrustList:

• Eka Robotics raises $150m to teach robot hands physics, not imitation

Sources

• Physicl: Physical AI fails in the long tail. We're building every edge case., 6 November 2026
• Physicl: About Physicl (read 9 November 2026), 9 November 2026