Physicl Raises $10M to Close Robotics’ Sim-to-Real Gap

Introduction

Robots rarely operate in the clean conditions represented by standard training examples. A cup may be hidden among other objects, placed in shadow, or presented in a setting the model has never seen before. Physicl is tackling that problem through more realistic digital environments, with the goal of narrowing the long-standing gap between simulation and real-world robot performance. The startup has raised $10 million in seed funding led by Inovia Capital.

From e-commerce rendering to physical AI

Physicl was founded by Alex de Vigan, who spent the previous decade building nfinite. That company helped brands and retailers create 3D models of products and display them online. As generative image models changed the e-commerce rendering market, de Vigan saw that the team’s underlying capabilities could be applied elsewhere: helping AI understand, interact with, and manipulate objects in space.

Training robots on video can produce limited returns, while collecting demonstrations through remote operation is expensive. Simulation offers a potential alternative, but only when the virtual world is sufficiently faithful to reality. Differences in geometry, materials, lighting, and physical behavior can cause a model that performs well in simulation to fail after deployment.

Physicl’s approach is to use its existing pipeline to transform incomplete and heterogeneous real-world information into detailed 3D representations. Those representations can then be used to create large numbers of synthetic training scenarios.

What the platform does

• Studio provides reusable building blocks: The product currently includes 10,000 3D environments and 100,000 assets representing different objects.
• It targets hard cases: Developers can create examples such as a cup on a cluttered countertop or a cup partly obscured by shadow.
• It aims to support an ongoing loop: Physicl does not want to be only a one-time data supplier. It envisions tools that generate new scenarios as a model’s training process reveals specific weaknesses.
• SpAItial is an early customer: The world-model company generates 3D simulations from 2D imagery and worked with Physicl to improve the fidelity of the synthetic data used by its systems, including Echo-3.

Why it matters

Physicl is entering a crowded and still loosely defined market. Startups are competing across data collection, labeling, synthetic-data generation, and software for foundation-model and robot developers. The company’s stated focus is post-training, particularly the edge cases that remain after a model has learned common situations.

That focus could be important because adding more ordinary examples does not necessarily solve a model’s most consequential failures. If a system can identify a missing type of scene from its training results and generate targeted variations, synthetic data could become part of a closed training loop rather than a static library assembled in advance.

The idea still faces substantial technical questions. Simulations must represent real physics accurately, generated examples must be useful rather than merely diverse, and performance must transfer reliably from the digital environment to hardware. Physicl’s funding gives it room to develop this tooling, while its origins in commercial 3D rendering provide a distinct starting point in the effort to make virtual worlds useful for physical AI.

Source: TechCrunch AI