Physicl raises $10 million to generate training data for robots’ hard cases

Physicl has raised $10 million to help robot and AI developers generate training examples of situations their models struggle with. The startup announced the seed round on November 6, 2026, alongside Studio, a tool for creating controlled 3D scenes rather than waiting to collect enough examples of difficult conditions in the real world.

Inovia Capital led the financing, with US Venture Partners participating. Physicl says the money will fund team growth and further development of its platform. Its focus is physical AI: models intended to understand or interact with physical environments, including systems used in robotics.

From retail renderings to robot training

Founder and CEO Alex de Vigan previously built nfinite, which helped retailers create and display 3D versions of their products. As generative AI pushed that business toward generated imagery, he saw another use for its technology. Physicl is applying the team’s experience in rendering environments to training problems faced by robot builders and developers of world models, TechCrunch reported.

The challenge goes beyond producing convincing pictures. Robot developers face expensive data collection through remote operation and disappointing returns from training on video, according to TechCrunch. Simulations offer a way to create training examples, but making those worlds faithful to real physics remains difficult. De Vigan argues that his team’s work on geometry, materials and physics can help address that problem.

Recreating the failures

Studio’s starting library includes 10,000 3D environments, more than 100,000 assets and billions of possible scene variations, Physicl says. The environments come with ground truth: precise reference information against which researchers can train and evaluate a model. The company’s pitch is not simply more data, but data chosen to address a particular weakness.

Researchers can select an environment or describe a scene where their model struggles. They can then change lighting, camera position and object placement to reproduce the conditions. Studio generates synthetic data—computer-created training examples—focused on that failure. Physicl positions this as post-training, meaning additional training aimed at weaknesses found after a model’s initial development.

De Vigan’s example is a home robot that recognizes a cup on a clear counter but misses it when clutter blocks part of it or shadows change. Studio is designed to recreate those conditions and generate variations. Researchers can use the resulting data to train against the failure, then test whether the model improves.

An impressive demo doesn’t tell you how a model will perform in all the situations it could encounter

Alex de Vigan, founder and CEO of Physicl

A customer result—and a bigger ambition

An early customer, SpAItial, builds world models that generate 3D environments from 2D imagery. Physicl says adding data generated with its library more than halved a SpAItial model’s error rate when reconstructing real rooms. That is a company-reported result about room reconstruction, not a measured improvement in a robot’s ability to handle objects.

SpAItial CEO Matthias Niessner told TechCrunch that manually curating data creates a long development cycle. He described the appeal of generating needed examples on demand, based on a model’s training state. That would make data generation part of ongoing development rather than a dataset assembled once and handed over.

De Vigan’s longer-term ambition is to let customers bring vast collections of images or lidar captures—measurements gathered using laser sensors—and turn them into 3D representations with many training variations. He presented that scale as a vision, not an achieved result. The seed funding now backs development of a platform intended to keep producing targeted data, rather than operate only as a supplier of fixed datasets.