Alex de Vigan’s last business venture was marred by the rise of generative AI, but he’s now betting that the technology his team built – a platform for creating compelling product renderings for e-commerce – can be repurposed to make robots more useful.
His new venture, Physicl, has raised $10 million in seed funding from Inovia Capital to help the world’s model designers and robot builders solve difficult training problems with more realistic simulations.
Founder and CEO de Vigan has spent the past decade building nfinite, which allows brands and retailers to easily create 3D models of their products and display them online. While the advent of transformer-based AI models led the company to focus more on generated images, de Vigan found new purpose in his team and the technology stack they built to faithfully render 3D environments.
“The very unique things we do in 3D were going to be replaced by AI, but AI was missing something: the ability to understand, interact with, and manipulate space and objects,” de Vigan says.
Researchers developing physical AI models to help robots perform more tasks, such as handling a variety of small objects, are frustrated by the benefits of training robots with video and the cost of collecting data remotely.
Although simulations in the digital world are seen as a promising solution, ensuring that these simulated worlds are faithful to real-world physics has proven difficult. de Vigan believes the pipeline his previous company built to “transform unstructured, incomplete, and heterogeneous real-world information into high-fidelity physical representations with the correct geometry, materials, and physics” can be used to more effectively create physical AI models.
Physicl joins a crowded category of startups competing to provide data and tools to basic model builders and robot operators. “This space is very new and very blurry at times,” de Vigan admits. Different companies are collecting, labeling, and generating data in different ways as customers seek to discover the secret sauce that allows robots to deliver on their promises.
In particular, Phyiscl wants to focus on post-training models in difficult edge cases. The startup has released a product called Studio, which has 10,000 3D environments and 100,000 assets representing a variety of objects. This allows users to create scenes and generate synthetic data to train robots on specific problems. For example, you can use Studio to generate training data to help a robot that has difficulty identifying a cup on a cluttered countertop or when the cup is in the shadow.
SpAItial, a global model builder, collaborated with Physicl as an early customer to improve product fidelity, and its latest product, Echo-3, was released last week. SpAItial’s tools generate 3D simulations from 2D images, but to do so effectively they need to be trained on high-fidelity synthetic data like the one provided by Physicl, said CEO and founder Matthias Niessner.
However, Physicl wants to be more than a data provider. As such, it is considered a tool for continuous synthetic data generation. “What we unknowingly created 10 years ago was an agent workflow that basically transformed unstructured information into high-fidelity 3D,” de Vigan says. He envisions customers bringing in trillions of images and millions of hours of LIDAR capture data and feeding it into the platform, which can be used to transform them into high-quality 3D representations and generate billions of permutations to train models.
“The real challenge is that you have to manually curate the data, which takes a very long cycle,” Niesner, who is also a professor of robotics at the Technical University of Munich, told TechCrunch. “What looks really appealing to us is that we can do this on the fly (…) When we run Claude, it checks, ‘Oh, we need more data about this for training,’ and it automatically generates that data depending on the state of the training. This is very exciting to me, and this is exactly what Physicl is offering.”
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
