Physicl raises $10 million to build simulations for robot training

Startup Physicl has raised $10 million in seed funding from Inovia Capital. The company develops tools and realistic 3D environments for training robots: its Studio product already offers 10,000 environments and 100,000 objects. Physicl founder Alex de Vigan wants to help developers of physical AI models tackle rare, complex situations that are difficult for robots to learn from video, while collecting data through teleoperation is expensive.

Physicl raises $10 million to build simulations for robot training

From 3D catalogs to robotics

For the past ten years, de Vigan has built nfinite, a company that helped brands and retailers create 3D product models and display them online. With the arrival of transformer-based AI models, nfinite shifted to image generation. De Vigan decided to apply the team’s accumulated expertise and technology to a different problem: AI still does not understand space and objects well enough to interact with and manipulate them.

Developers of physical AI models need to teach robots to work with a range of objects, including small ones. Training on video delivers insufficient results, while collecting data through teleoperation is costly.

Digital simulations can help if they accurately reproduce real-world physics. Physicl uses a pipeline developed to turn fragmented and incomplete information about the real world into accurate physical representations, with the right geometry, materials and physical properties.

Edge cases and synthetic data

There are already startups in the market that collect, label and create data in different formats for foundation-model developers and robot operators. De Vigan says the industry is still young, and its boundaries are not yet clearly defined.

Physicl is focused on fine-tuning models on rare, complex cases. The company has launched Studio, which includes 10,000 3D environments and 100,000 objects. Users can assemble scenes and generate synthetic data for specific tasks. For example, they can train a robot to spot a cup on a cluttered countertop or recognize one in shadow.

One of Physicl’s first customers was SpAItial, which develops world models. The company worked with the startup to improve the accuracy of its products. Its latest product, Echo-3, launched last week. SpAItial’s tool creates 3D simulations from 2D images; the company’s founder and CEO, Matthias Niessner, said the models needed to be trained on realistic synthetic data from Physicl to do this.

From data to automated generation

Physicl wants to be more than a data provider: it also aims to become a tool for generating data continuously. De Vigan expects customers will be able to upload trillions of images or millions of hours of lidar data to the platform, convert them into high-quality 3D representations and create billions of variations for training models.

Niessner, who is also a robotics professor at the Technical University of Munich, noted that manually selecting data takes a long time. In the future, a model like Claude could detect a shortage of training examples and automatically create more data based on the current state of training.

Dataist’s take: Physicl connects two stages of working with robots: preparing realistic simulations and generating data as models identify gaps in their training. For now, the company is describing an anticipated use case, not a proven scale of adoption: the source provides no data on how widely this approach is already being used.