Physicl Raises $10 Million Seed Round to Build Realistic Simulations for Training Robots

The company is founded and run by Alex de Vigan, who spent the past decade building nfinite. That business made 3D models of products so brands and retailers could show them online, and it moved towards generative imagery once transformer-based AI arrived. De Vigan told TechCrunch that generative AI was set to replace much of what his team did in 3D, but that AI still lacked an understanding of space and objects and how to handle them. He is now pointing the rendering pipeline nfinite built at that gap.

The bet targets one of the main bottlenecks in physical AI. Developers trying to teach robots to handle varied, small objects have had disappointing results from training on video, and collecting data by operating robots remotely is expensive. Simulated worlds are a cheaper alternative, but making them obey real-world physics closely enough to be useful has proved hard. De Vigan says the pipeline can take information that is "unstructured, incomplete, heterogeneous" and turn it into a representation with correct geometry, materials and physics.

Physicl is aiming at post-training, the stage where models are refined on difficult edge cases. Its first product, Studio, comes with 10,000 3D environments and 100,000 object assets. Customers use it to build scenes and generate synthetic data for specific failures, such as a robot that cannot pick out a cup on a cluttered counter, or a cup sitting in shadow.

An early customer is SpAItial, a world model developer that generates 3D simulations from 2D images and released its latest model, Echo-3, last week. Its founder and chief executive, Matthias Niessner, a professor at the Technical University of Munich, said his team needed high-fidelity synthetic data of the kind Physicl produces. He told TechCrunch that curating such data by hand is "a very long cycle", and that the appeal is being able to generate more of a particular kind of data automatically while a model is still training.

The start-up is entering a crowded field of companies that collect, label or generate data for robot makers and foundation model builders, and de Vigan acknowledged that the category is young and its boundaries unclear. He wants Physicl to be seen as ongoing tooling rather than a one-off data supplier. He described customers feeding in trillions of images or millions of hours of lidar capture and using the platform to produce billions of variations to train their models.