Physicl raises $10 million for realistic robot training simulations

Seed round and the founder's pivot

Startup Physicl has raised $10 million in seed funding from Inovia Capital, TechCrunch reports. Founder and CEO Alex de Vigan spent ten years building nfinite, a company that helped brands and retailers create 3D models of products for online storefronts. With the arrival of generative AI based on transformers, that business came under pressure, and de Vigan decided to redirect the team and the accumulated technology stack toward a new task — realistic simulations for training robots. According to him, AI was supposed to replace what they were doing in 3D, but it lacked the 'ability to understand, interact with, and manipulate space and objects.'

Why robots need digital worlds

Developers of 'physical AI' that trains robots to manipulate a variety of small objects face two problems: training on video yields weak returns, and collecting data through teleoperating robots is expensive. Simulations in digital worlds are considered a promising solution, but achieving their accurate match to real-world physics is difficult. De Vigan believes that the nfinite pipeline, which turned 'unstructured, incomplete, heterogeneous information from the real world into a high-fidelity physical representation with correct geometry, materials, and physics,' can be applied precisely to this task.

The Studio product and first customers

Physicl is entering an already crowded market of startups supplying data and tools to creators of foundation models and robot operators; de Vigan himself admits that 'the space is very young and sometimes very blurry.' The company has released a product called Studio with 10,000 3D environments and 100,000 asset objects: users can assemble scenes and generate synthetic data to work through specific problems, such as recognizing a cup on a cluttered countertop or in shadow. The focus is on post-training models on difficult edge cases. An early customer was world-model developer SpAItial, whose product Echo-3, which generates 3D simulations from 2D images, launched last week; its founder Matthias Niessner, also a professor of robotics at the Technical University of Munich, said they needed exactly this kind of high-fidelity synthetic data.

Ambitions beyond data supply

Physicl sees itself not merely as a data supplier but as a tool for continuous generation of synthetic data: de Vigan says that ten years ago, without realizing it, he created 'an agentic workflow for turning unstructured information into high-fidelity 3D.' He envisions customers uploading trillions of images or millions of hours of LIDAR data, which the platform turns into high-quality 3D representations for generating billions of variations during model training. Niessner adds that manual data curation is 'a very long cycle,' and Physicl's appeal lies in being able to do it 'on the fly': the training system itself determines what data is missing and automatically generates it.