Physicl raises $10M seed to build the scenes robot models fail in

Physicl has raised a $10 million seed round led by Inovia Capital, with US Venture Partners participating, and launched its first product, Studio, on Nov. 6. A researcher describes a case their model gets wrong, such as a mug half hidden behind a bag at sunset. Studio picks a matching scene from 10,000 environments and 100,000 assets built from real 3D geometry, renders it thousands of ways, and computes depth, surface normals, segmentation masks, boxes and camera pose from the scene itself, the company says.

What Studio ships today is perception data. It renders images and video, not robot actions or contact forces, so the teams it helps first are those whose models fail at seeing. The one result Physicl names fits that: SpAItial, which turns real rooms into explorable 3D, more than halved its reconstruction error on real rooms with Physicl data. SpAItial's chief executive, Matthias Niessner, is also a Physicl advisor.

Chief executive Alex de Vigan spent the past decade at nfinite building 3D models of products and rooms for retailers, and Physicl reuses that pipeline, TechCrunch reports. The round funds teams in Paris and the San Francisco Bay Area. Studio is in free beta, opening to research teams a few at a time.

Physicl says robotics labs were among the first teams to ask for its data. A robot team publishing a real-world success rate before and after training on Studio scenes is the result its physical AI pitch still needs; recent bets on robot data, such as LiberAI's glove recordings, went after actions instead.