The previous startup was impacted by the rise of generative AI, but now they are betting on repurposing the technology developed by that team back then—a platform that creates attractive product renderings for e-commerce—to enhance the practicality of robots.
His new company, Physicl, has received a $10 million seed round of financing from Inovia Capital with the aim of helping world model designers and robot manufacturers to solve challenging training problems with more realistic simulations.
The founder, along with CEO, de, and Vigan, has been working on nfinite for the past decade. This company enables brands and retailers to easily create 3D models of their products and display them online. With the emergence of AI models based on Transformer, the company has begun to focus more on generative imaging. However, de and Vigan have seen new uses for the team and their technology stack: high-fidelity rendering of 3D environments.
de Vigan represents: "The part where we are very unique in the 3D field. AI would have replaced it, but AI lacks something: its ability to understand, interact with, and manipulate space and objects."
Researchers have been developing physical AI models in the hope of making robots more adept at handling tasks involving a variety of small objects, but they have been disappointed with the results achieved through video training alone. They also believe that the cost of collecting data via remote operation is too high.
Simulating in the digital world is regarded as a promising solution, but it turns out to be not easy to ensure that these simulated worlds are faithful to the physical laws of the real world. de Vigan believes that the process established by his previous company – which involves transforming unstructured, incomplete, and heterogeneous real-world information into high-fidelity physical representations… with the correct geometry, materials, and physical properties – can now be used to improve the effectiveness of physical AI models.
Physicl Entering this field is like stepping into a crowded racecourse, where many startups are vying for the opportunity to provide data and tools to developers of foundational models and operators of robots. de Vigan also acknowledges, "This field is still very young, and sometimes the boundaries are quite blurred." Currently, various companies are collecting, annotating, or generating data in different modalities, while their clients are still figuring out what "secret formula" will truly enable robots to fulfill their promises.
Physicl particularly hopes to focus on the difficult edge cases during the post-training phase of the model. This startup is launching a product called Studio, which includes 10,000 3D environments and 100,000 assets representing different objects. Users can use this to create scenarios and generate synthetic data for training robots to solve specific problems. For example, Studio can generate training data to help robots identify cups on a cluttered table, or to be able to distinguish them even when they are in shadows.
World model developer SpAItial collaborated with Physicl as an early customer to enhance the fidelity of their products. Echo-3, the latest product from SpAItial, was released last week. The tools from SpAItial can generate 3D simulations from 2D images. CEO, one of the co-founders of SpAItial, stated that to achieve this effectively, they need to train using the high-fidelity synthetic data provided by Physicl.
However, Physicl hopes to be more than just a data provider; it positions itself as a tool platform that continuously generates synthetic data. de and Vigan said, "What we created 10 years ago was, in essence, a type of proxy workflow that could transform unstructured information into high-fidelity 3D content." They envision that customers could input trillions of images or millions of hours of LIDAR collected data into their platform, convert it into high-quality 3D representations, and then use these to generate billions of different combinations for training models.
Niessner is also a professor of robotics at the Technical University of Munich. He said about TechCrunch: "The real challenge lies in the fact that you have to manually organize the data, and that is a very long process. What is very attractive for us is that you can complete this task instantly... When you run Claude, it will check and say, 'Oh, I need more of this kind of data for training,' and then it will automatically generate this data based on the training status. In my opinion, that is very exciting, and that is exactly what Physicl offers."
