Physicl has raised $10 million in seed funding from Inovia Capital to help robot builders train models in digital environments that better reflect the physical world. The company’s Studio product lets users assemble scenes and generate synthetic data for specific training challenges, including recognizing objects in clutter or shadow.
From product renderings to robot training
Founder and CEO Alex de Vigan spent the last decade building nfinite, a company that helped brands and retailers create 3D models of products for online display. As transformer-based AI led nfinite toward generative imagery, de Vigan saw an opportunity to apply the team’s rendering technology elsewhere.
He argues that AI’s growing ability to generate images does not solve the challenge of understanding and interacting with space and objects. The 3D pipeline developed at his previous company, he believes, can help create more useful representations for physical AI systems.
That pipeline was built to turn incomplete and varied real-world information into detailed 3D scenes with geometry, materials, and physics. Physicl is adapting it to make simulated settings more faithful to the environments where robots need to operate.
Simulations target costly training gaps
Researchers working on physical AI models want robots to handle tasks involving varied, small objects. The source describes two difficulties: training on video has produced frustrating results, while collecting data through remote operations can be expensive.
Simulation offers another route. But digital environments only help if they represent real-world physics well enough for a model to learn from them. A scene that lacks accurate objects or conditions may leave a robot unprepared for the situation it is meant to handle.
Physicl is focusing on post-training, especially difficult edge cases. Studio includes 10,000 3D environments and 100,000 assets representing different objects. Users can build scenes around specific problems, such as a cup that is hard to distinguish on a cluttered counter or when it sits in shadow.
A tool for generating synthetic data
The company’s ambition goes beyond supplying a fixed dataset. It wants to provide tools that let customers create synthetic training data as their needs change. De Vigan describes a workflow that could take large collections of images or LIDAR captures, turn them into 3D representations, and generate many variations for model training.
That approach could address the time required to curate training data by hand. Matthias Niessner, CEO and founder of world model builder SpAItial, told TechCrunch that on-demand generation is appealing because it could produce more data in response to a model’s training state.
In the workflow Niessner described, a model could identify a gap in its training and prompt the system to generate examples for it. Physicl presents this kind of adaptive synthetic data generation as a core part of what it offers.
Early collaboration in a crowded field
SpAItial worked with Physicl as an early customer to improve the fidelity of its products. Its tool generates 3D simulations from 2D imagery, and Niessner said the team needed high-fidelity synthetic data to train it effectively. The company’s latest product, Echo-3, was released last week.
Physicl is entering a young, crowded market where startups collect, label, or generate different kinds of data and tooling for foundation model builders and robot operators. The source describes the category as still taking shape, as customers work out what will help robots perform as promised.
For Physicl, the bet is that realistic 3D scenes can make robot training more targeted and easier to update. If synthetic environments can be generated around the gaps a model encounters, builders may be able to address specific failures without relying only on video or costly remote data collection.