Physicl Raises $10M Seed Round to Build Realistic Training Worlds for Robots
Physicl, a startup building 3D simulation data for robots and other physical AI systems, has raised $10 million in seed funding. Inovia Capital led the round, with participation from US Venture Partners. The company announced the raise on November 6, 2026, alongside the beta launch of its first product, Studio.
The funding lands at a moment when robotics teams are running into the same wall: models that look strong in the lab often fail in messy real-world settings. Physicl is betting that better synthetic data, built to match real-world physics, is the fastest way through that wall.
What Physicl does
Physicl sells physics-aware 3D training data and tooling to two groups: companies building foundation models for robots and world models, and teams operating robots in the field. Its goal is to make simulated training environments faithful enough that what a robot learns in simulation carries over to the real world.
The company focuses on the hard edge cases that trip up physical AI. These include occluded objects, shifting shadows, cluttered surfaces and items left in unexpected places. These scenarios are rare in real-world datasets, expensive to capture with cameras, and exactly where robots tend to fail.
Rather than competing to train its own robot models, Physicl positions itself as a data infrastructure layer. It supplies the environments, assets and labels that model builders use for training and, increasingly, for post-training on difficult cases.
Inside Studio, Physicl's first product
Studio launched in beta alongside the funding announcement. It lets researchers describe a scenario in plain language, generate a matching 3D scene, adjust the conditions and export synthetic images or video for training.
At launch, Studio includes:
10,000 3D environments covering the indoor and real-world settings robots operate in
100,000 3D assets that can be placed, swapped and rearranged within those scenes
Adjustable conditions, including lighting, objects, materials and camera positions
Training labels exported with each scene, such as depth maps, object segmentation and camera positions
That last point matters as much as the visuals. Labelling real-world footage by hand is slow and costly, and errors in labels flow straight into model errors. A synthetic scene comes with its ground truth built in, because the system knows exactly where every object and surface sits.
Studio is opening gradually to research labs, universities and companies building robotics and AI models.
Early results with SpAItial
Physicl's first named customer is SpAItial, a company building world models, led by CEO Matthias Niessner, a professor at the Technical University of Munich. According to Physicl, SpAItial cut its 3D reconstruction error on real-world rooms by more than half after training on Physicl's synthetic data. SpAItial led that evaluation.
This is an early, company-reported result rather than an independent benchmark. Still, it points to the core claim behind Physicl's pitch: well-built synthetic data can measurably improve how models perform on real-world scenes, not just simulated ones.
The team behind Physicl
Physicl is led by founder and CEO Alex de Vigan. He spent roughly a decade building nfinite, a company that produced realistic 3D product visuals for brands and online retailers. That experience in producing large volumes of photorealistic 3D content is the foundation Physicl now applies to robotics.
The company has about 30 employees across Paris, San Francisco and New York. The team combines expertise in 3D technology, simulation and machine learning.
Why physical AI needs better training data
Large language models were trained on a vast supply of text scraped from the internet. Robots have no equivalent. There is no internet-scale archive of robots grasping, navigating and recovering from mistakes across millions of real homes, warehouses and factories.
Teams have three main ways to close that gap, and each has limits:
Real-world data collection is accurate but slow, costly and hard to scale, especially for rare events.
Teleoperation and demonstrations give high-quality examples but require people to control robots for many hours.
Simulation and synthetic data scale cheaply, but only help if the simulated world behaves like the real one.
The third approach is where Physicl plays. Its bet is that the "sim-to-real gap", the drop in performance when a model moves from simulation to reality, can be narrowed with physics-aware assets and carefully varied scenes. If that holds, synthetic data becomes a practical way to cover the long tail of edge cases that real-world collection misses.
How Physicl will use the funding
Physicl plans to use the $10 million to expand its team and continue building its platform for generating synthetic training data. The platform targets robotics, autonomous systems and physical AI models more broadly.
The company has not disclosed its valuation, pricing or revenue.
What to watch next
Three questions will shape how far Physicl goes from here:
Independent results: further customer evaluations, beyond SpAItial's, will show whether the reported gains hold across other models and tasks.
Customer mix: whether demand comes mainly from foundation-model labs or from companies deploying robots will shape the product roadmap.
Studio's rollout: how quickly the beta opens to more labs and companies will signal how ready the platform is for production use.
For now, the round gives Physicl capital and a launched product in one of the fastest-moving corners of AI. As more companies try to bring robots out of controlled environments and into the real world, the quality of their training data is likely to decide which ones succeed.
Key facts
Company: Physicl
Round: Seed
Amount: $10 million
Lead investor: Inovia Capital
Other investors: US Venture Partners
Announced: November 6, 2026
CEO: Alex de Vigan
Offices: Paris, San Francisco, New York
Product: Studio (beta)
Sources: Dealroom, Global Startups Insights, piQ Markets