Founded by former nfinite CEO Alex de Vigan, the Paris startup is developing synthetic training data for robots and world models that struggle with unusual real-world situations.
Physicl has raised $10 million in seed funding led by Inovia Capital, with participation from U.S. Venture Partners.
The financing, announced November 6, will support the Paris-based company's development of realistic synthetic training data for physical artificial intelligence.
Founded by Alex de Vigan, Physicl is building infrastructure that helps AI researchers train models to understand and interact with three-dimensional environments.
The startup is targeting one of robotics' biggest technical challenges: the gap between impressive laboratory demonstrations and reliable performance in unpredictable real-world conditions.
Training Models on Their Weaknesses
AI models can perform well on familiar tasks but fail when they encounter unusual lighting, cluttered environments or unexpected object positions.
A household robot may recognize a cup on an empty table while struggling to identify the same object when it is partially hidden.
Collecting enough real-world examples of every possible situation is expensive and time-consuming.
Physicl allows researchers to reproduce those scenarios digitally and generate training data designed around specific model weaknesses.
Its first product, Studio, provides access to 10,000 3D environments and more than 100,000 assets.
Researchers can modify lighting, camera angles, geometry and object placement, creating large numbers of controlled training examples.
The company emphasizes ground-truth data, allowing researchers to evaluate whether models understand a scene correctly.
From E-Commerce to Robotics
De Vigan previously founded nfinite, a company that developed 3D visualization infrastructure for retail and e-commerce.
The technical systems used to reconstruct products and environments are now being adapted for physical AI.
Physicl reports working with organizations including Meta, World Labs and SpAItial.
In one collaboration, SpAItial reported reducing reconstruction errors by more than half after incorporating Physicl-generated training data.
The result is promising but should be treated as a reported collaboration outcome, not proof of equivalent performance across all robotics tasks.
The Business Opportunity
As more companies develop robotics foundation models, synthetic data could become a major infrastructure market.
Instead of collecting every training example through physical robots, developers may increasingly generate targeted situations in simulation.
But those simulations must accurately represent real-world geometry, materials and physics.
The challenge for Physicl is demonstrating that models trained on its synthetic environments consistently perform better outside the simulation.
The $10 million financing gives the company resources to expand its product and compete in a rapidly developing market for physical AI infrastructure.
Sources: GlobeNewswire · TechCrunch · The Next Web