Physicl raises $10M to build realistic 3D simulations for robot training

From e-commerce to robotics

Alex de Vigan is pivoting from the world of online shopping to the frontier of physical AI. His previous company, nfinite, specialized in helping retailers create high-quality 3D models of products to display on websites. While generative AI eventually shifted that business toward imagery, de Vigan realized the underlying technology had a different potential.

The core technology involved creating faithful 3D representations of objects from imperfect real-world information. De Vigan noticed that while current AI models are great at generating pictures, they often struggle to understand how objects actually occupy and interact with physical space. He believes this gap is the primary hurdle for making robots truly useful in human environments.

By repurposing his team's tech stack, Physicl aims to solve the 'physics' problem for robot builders. The goal is to move beyond simple visuals to create digital spaces where materials, geometry, and physics are all accurate. This allows robots to practice complex movements in a virtual world before they ever attempt them in reality.

The problem of synthetic data

Researchers currently face a difficult choice when training robots to perform tasks like handling small, diverse objects. They can either use video data, which often lacks the depth needed for physical interaction, or they can use remote operations, which are incredibly expensive and slow to scale. This creates a bottleneck for companies trying to build reliable autonomous systems.

Simulations offer a middle ground, but only if those simulations are realistic enough to be useful. If a robot learns to pick up a cup in a digital world where the cup doesn't have the correct weight or friction, it will fail when it tries the same task in a real kitchen. Ensuring these digital worlds are faithful to real-world physics has proven to be a major technical challenge for the industry.

Physicl intends to address this by providing a pipeline that transforms unstructured data into high-fidelity representations. This includes taking things like LIDAR captures or millions of images and turning them into 3D environments. The startup wants to provide the tools that allow developers to generate billions of different permutations to train their models effectively.

How Studio works

To help developers get started, Physicl is releasing a product called Studio. This platform includes 10,000 3D environments and 100,000 assets representing various objects. It is designed to let users create specific scenes to generate synthetic data for training robots on very particular problems.

For example, a developer could use Studio to create a scene of a cluttered countertop with a cup hidden in shadows. The system can then generate thousands of variations of that scene to help a robot learn to discern the object. This type of targeted training is essential for handling the 'edge cases' that often cause robots to fail in unpredictable real-world settings.

The company describes its work as an agentic workflow for 3D transformation. Instead of humans manually building every single room or object, the platform is designed to handle the heavy lifting of turning raw data into usable 3D assets. This allows for a faster cycle of data generation as models become more sophisticated.

Early adoption and scale

Physicl is entering a crowded market where many startups are trying to provide data and tooling for foundation model builders. The space is still very young, and companies are experimenting with different ways to label, collect, and generate data. Physicl is positioning itself as a primary tool for ongoing synthetic data generation rather than just a one-time data provider.

The startup already has early customers, including SpAItial, a world model builder. SpAItial collaborated with Physicl to improve the fidelity of its products, such as the recently released Echo-3. Because SpAItial generates 3D simulations from 2D imagery, it required the high-fidelity synthetic data that Physicl's pipeline provides to be effective.

Matthias Niessner, the founder of SpAItial and a robotics professor, noted the importance of this automated approach. He explained that while curating data by hand is a long cycle, Physicl allows for data to be generated on the fly. This means a model could potentially identify a need for more data and have the platform auto-generate it based on the current training state.