Generative AI may have upended many traditional software businesses, but one entrepreneur is repurposing the 3D-rendering stack from his previous company to tackle one of robotics' hardest training problems. Alex de Vigan, whose prior venture was impacted by the rapid rise of transformer models, is now betting that lifelike digital environments can bridge the gap between artificial intelligence and the physical world.
Securing Seed Funding for Robotics
His new venture, Physicl, has successfully raised $10 million in seed funding from Inovia Capital. The capital will be deployed to help world model designers and robot builders overcome difficult training bottlenecks by supplying them with highly realistic simulations that mirror real-world physics.
Pivoting from E-Commerce to Robotics
Founder and chief executive officer de Vigan spent the past ten years building nfinite, a platform that enabled brands and retailers to generate 3D models of their merchandise for online retail. While the arrival of advanced AI models shifted that company's focus toward generative imagery, de Vigan recognized an entirely new application for the underlying technology stack his team had engineered to accurately render complex 3D environments.
Addressing the Limitations of AI
De Vigan noted that while AI was rapidly evolving to replace many traditional 3D workflows, artificial intelligence still struggled with a fundamental limitation regarding its ability to understand, interact with, and manipulate physical space and everyday objects. Researchers developing physical AI models to give robots better dexterity with small, diverse items have faced persistent frustration when relying solely on video training or expensive remote operation data collection.
Building Faithful Digital Simulations
While digital simulations have long been viewed as a promising alternative, ensuring those virtual worlds strictly adhere to real-world physics has proven exceedingly difficult. De Vigan believes the engineering pipeline built by his previous company to transform unstructured, heterogeneous real-world data into high-fidelity physical representations with correct geometry, materials, and physics can now make physical AI models significantly more effective.
Launching Studio for Synthetic Data
Physicl enters a crowded ecosystem of startups providing tooling and data solutions to foundation model developers and robot operators. To carve out its niche, the startup is launching a product called Studio, which features 10,000 3D environments and 100,000 assets representing various everyday objects. The tool allows users to generate synthetic training scenes for edge cases, such as helping a robotic arm identify a cup resting on a cluttered countertop or sitting in deep shadow.
Collaboration with SpAItial
World model builder SpAItial collaborated with Physicl as an early customer to enhance the fidelity of its simulation pipelines. SpAItial recently released Echo-3, its latest tool that generates 3D simulations from 2D imagery, which required training on the high-fidelity synthetic data supplied by Physicl. SpAItial CEO and founder Matthias Niessner emphasized that accessing realistic synthetic data is crucial for advancing spatial intelligence models.
Scaling Up Ongoing Data Generation
Beyond acting as a static data provider, Physicl aims to serve as the foundational tooling for continuous synthetic data generation. De Vigan envisions customers feeding trillions of images or millions of hours of LiDAR capture data into the platform to generate billions of training permutations. Niessner praised the dynamic nature of the workflow, noting that automated generation based on real-time training needs represents a major leap forward for the robotics industry.
