Physicl raises $10 million for robot-training simulations

Researchers have struggled to teach robots to handle varied objects using video alone, while collecting training data through remote operation can be expensive. Simulations offer an alternative, but they must reproduce real-world geometry, materials and physics accurately. De Vigan says technology developed at nfinite can turn incomplete, varied real-world information into high-fidelity 3D representations.

Physicl is launching Studio, a product with 10,000 3D environments and 100,000 assets. Users can build scenes and generate synthetic training data for specific challenges, such as helping a robot identify a cup on a cluttered counter or in shadow. The company is targeting post-training on difficult edge cases, in a young and crowded market where startups are supplying data and tools to model developers and robot operators.

SpAItial, a world-model company and early Physicl customer, worked with the startup to improve its simulations. Its latest product, Echo-3, was released last week. SpAItial founder and CEO Matthias Niessner said the company needed high-fidelity synthetic data to train its tool, which creates 3D simulations from 2D images.

Physicl aims to provide tools for generating synthetic data continuously, not just act as a data supplier. De Vigan described a system that could turn large collections of images or LiDAR captures into 3D representations, then produce variations for model training. Niessner said automatic generation could reduce the lengthy process of manually curating data and allow systems to create more examples as training needs emerge.