Physicl raises $10 million to help robots learn in virtual environments

Alex de Vigan, whose previous business was impacted by the emergence of generative AI, is now utilizing his team's existing technology to improve robotics. His new startup, Physicl, has successfully raised $10 million in seed funding, with the investment round led by Inovia Capital. The capital will be used to assist designers of world models and robot developers in overcoming complex training hurdles through the use of highly realistic simulations.

During the last ten years, CEO and founder de Vigan led nfinite, a company that helped retailers and brands produce 3D models of products for online display. Although the rise of transformer-based AI models caused nfinite to shift its focus toward generative imagery, de Vigan identified a new application for the technical infrastructure his team had developed to render 3D environments accurately.

"What we were super unique about doing in 3D, AI was going to replace, but something was missing in AI: its ability to understand, interact, and manipulate space and objects," de Vigan stated. Developers of physical AI models, who aim to improve a robot's ability to handle various small items, have encountered difficulties with training via video and the high costs associated with remote data collection.

Digital simulations are considered a potential fix, though creating environments that accurately reflect real-world physics remains a challenge. De Vigan maintains that the system his former company created to "transform unstructured, incomplete, heterogeneous real-world information into a high-fidelity physical representation… with correct geometry, materials, and physics" can now enhance the effectiveness of physical AI models.

Physicl is joining a busy sector of startups that aim to provide tools and data to foundation model creators and robotics operators. "The space is very young and very blurry sometimes," de Vigan acknowledged. A range of firms are currently gathering, tagging, or producing data in various forms as they attempt to discover the methods necessary to help robots reach their full potential.

Physicl intends to concentrate on post-training models for complex edge cases. The company is launching a tool named Studio, which includes 100,000 assets and 10,000 3D environments. This platform allows users to build scenes to produce synthetic data for specific robotics training, such as helping a machine identify a cup on a cluttered counter or when the object is obscured by shadows.

World model creator SpAItial worked with Physicl as an early client to boost the quality of its offerings, including the recently released Echo-3. SpAItial CEO and founder Matthias Niessner explained that their tool, which creates 3D simulations from 2D images, required training on the high-fidelity synthetic data that Physicl provides.

Physicl aims to function as more than just a data supplier, positioning itself as a platform for continuous synthetic data generation. "What we created 10 years ago, without knowing it, was basically an agentic workflow to transform unstructured information into high-fidelity 3D," de Vigan remarked. He anticipates clients will use the platform to process trillions of images or millions of hours of lidar data into 3D representations, creating billions of variations for model training.

"The real challenge is, you have to curate the data by hand, and this is a very long cycle," Niessner, who also serves as a robotics professor at the Technical University of Munich, said. "To us what seems very attractive is you can just do this on the fly […] You’re running Claude, and it checks, ‘Oh, I need more data of this in my training,’ and then it auto generates that data, depending on the training state. That, to me, is super exciting, and this is exactly what Physicl is offering."