AI
Physicl nabs $10M to turn photorealistic 3D simulations into robot training data
A startup led by an e-commerce 3D graphics veteran is pivoting to help roboticists create faithful digital environments for training physical AI models.
TechCrunch — AI
Alex de Vigan's previous company built tools for rendering high-fidelity 3D product models in online retail—until generative AI models made those pipelines less essential. Now he's backing a new bet: that the same rendering expertise can power more capable robots by generating hyper-realistic training data in controlled digital settings.
Physicl, de Vigan's latest venture, has secured $10 million in seed funding from Inovia Capital. The platform targets a persistent headache in robotics research: training physical AI systems to handle real-world tasks cheaply and at scale.
The simulation problem#
Roboticists building models for manipulation tasks—grasping objects, navigating clutter, handling edge cases—have struggled with traditional training approaches. Video-based learning produces inconsistent results, while collecting real-world data through teleoperation is prohibitively expensive and time-consuming.
Digital simulations offer a way out, but only if the virtual world accurately mirrors physical reality. Creating high-fidelity 3D environments with correct geometry, materials, and physics has proven to be the technical bottleneck. Physicl's core insight is that the pipeline de Vigan built over a decade at his prior company, nfinite, for faithfully converting raw visual information into precise 3D representations, can solve this problem for roboticists.
A crowded but nascent market#
The startup is entering a fast-growing but still-forming ecosystem of companies building data and tools for foundation models and robot operators. de Vigan acknowledges the space remains "very young and very blurry sometimes," with competitors collecting, labeling, or generating data across various formats as customers search for the approaches that will unlock robot capabilities.
Physicl's immediate focus is on post-training—helping models learn to handle hard cases and outliers. The company is launching a product called Studio, featuring a library of 10,000 pre-built 3D environments and 100,000 configurable assets. Users can compose scenes to generate synthetic training data targeting specific robotic challenges—for instance, teaching a gripper to recognize a cup on a messy counter or under shadow.
Early validation and ambition#
Early customer SpAItial, a world-model builder, used Physicl's data to improve its latest product, , released this week. SpAItial's technology generates 3D simulations from 2D images; Matthias Niessner, the company's CEO and a robotics professor at the Technical University of Munich, noted that training on Physicl's high-fidelity synthetic data was essential to achieving better results.

