Physicl raises $10M seed to make simulations real for robots

Targeting the sim-to-real gap

Researchers building physical AI models for tasks such as handling small, varied objects have been frustrated by the results of training on video, and by the cost of collecting data through remote operation. Simulation is seen as a promising alternative, but keeping virtual worlds true to real-world physics has been hard.

De Vigan argues that the pipeline his earlier company built can turn unstructured, incomplete real-world information into a high-fidelity physical representation with correct geometry, materials and physics. Physicl is focusing on post-training models on difficult edge cases.

Its first product, Studio, ships with 10,000 3D environments and 100,000 object assets. Users build scenes to generate synthetic training data for specific problems, for example a robot that struggles to spot a cup on a cluttered counter or when the cup sits in shadow.

Early customer and competition

World model builder SpAItial worked with Physicl as an early customer. Its tool generates 3D simulations from 2D imagery, and its latest model, Echo-3, was released last week. CEO Matthias Niessner, also a robotics professor at the Technical University of Munich, said training such a tool requires high-fidelity synthetic data. He likes the idea of generating data on the fly, with an AI agent detecting what the training run lacks and requesting it automatically.

Physicl wants to be a tooling layer for continuous synthetic data generation rather than just a data supplier. De Vigan imagines customers feeding in trillions of images or millions of hours of lidar data and getting billions of permutations back. He admits the field is crowded: "The space is very young and very blurry sometimes."