Physicl Raises $10M Seed to Address Physical AI’s Edge-Case Problem

Physicl announced on November 6, 2026, that it had raised $10 million in seed funding led by Inovia Capital, with participation from US Venture Partners, and launched Studio, its first application for generating synthetic training data for physical AI.

The round funds the company’s team in Paris and the San Francisco Bay Area and the platform behind Studio. Inovia partner Steve Woods described Physicl’s scalable, high-quality physical data grounded in real-world physics as a critical part of translating model capabilities into real-world performance.

Physicl is targeting rare conditions that can determine whether robots, autonomous systems and world models operate reliably outside demonstrations, including a cup partly hidden by a bag, a shadow crossing a counter or a box stacked slightly out of position. These cases are difficult to capture at scale and, once recorded, cannot be replayed with selected conditions changed.

Physicl argues that physical AI lacks a training corpus comparable to the internet text behind language models or the billions of captioned photos behind image models, leaving teams short of examples for the failures that can block deployment.

Founder and CEO Alex de Vigan has spent ten years digitizing objects and rooms in 3D, including building libraries and variation software for large retailers. Robotics, world-model and vision teams subsequently began requesting precisely labeled versions of situations their models failed to handle, in volumes of thousands. That demand provided the basis for Physicl’s synthetic-data platform.

Physicl has roughly 30 team members across 3D, simulation and machine learning, with offices in New York, San Francisco and Paris. Most spent six years building 3D production systems with de Vigan before he founded the company.

Studio Generates Controlled Edge Cases

A product announcement published alongside the funding news describes a four-step workflow. A researcher describes a failure in plain language; Studio proposes a matching scene and variations; the researcher adjusts their ranges and fixes conditions such as the camera path; and the system renders and exports the result while recording the conditions behind each frame.

Studio draws on a library of 10,000 environments and 100,000 underlying assets built from real 3D geometry and known materials. Lighting, time of day, clutter, objects, materials, camera height and camera path can be changed independently or together. That permits controlled comparisons in which only selected conditions differ between clips.

Depth, surface normals, instance and class segmentation and two-dimensional boxes come from the same 3D scene that produces the image, rather than being predicted from the rendered picture or drawn by annotators. Exports include RGB images or video at resolutions up to 1920 by 1080, with per-frame camera pose and scene metadata covering lighting, layout and objects.

Beta Access and an Early Result

Studio is in beta and free during that period, with access opening in waves to laboratories, universities and companies developing vision, world-model and robotics models. Real-world data remains essential to the approach: Studio is designed to supplement it with rare cases that real capture cannot supply repeatedly and at volume. New scenes are added daily, and Physicl plans to develop the controls, ground-truth layers and export formats through feedback from research users.

Physicl reported that world-model company SpAItial more than halved its reconstruction error on real rooms after training with Physicl data. The evaluation came from Matthias Niessner’s team at TUM. Niessner is SpAItial’s CEO and a TUM professor, and is also listed as a Physicl advisor.

Studio’s beta puts Physicl’s premise to a wider test: whether repeatable, exactly labeled synthetic edge cases can improve models reliably enough for deployment in physical environments.