AI is transforming everything around us, but so far it has remained largely contained to the digital realm. But more and more startups are trying to bring this into the real world.
Perceptron is one of the startups founded by two former Meta researchers. Founded in November 2024, the company develops Frontier Vision models that aim to enable machines to better interact with their physical environment.
This week, the company released its latest model, Isaac 0.5. The model is designed to give machines the ability to “perceive, reason and act” in industrial environments, the developers said. Specifically, this software can enable visually guided robots to navigate complex environments such as warehouses and factory floors. It also helps businesses extract visual intelligence from videos recorded by these bots.
Isaac 0.5 is also released as an open weight model, so anyone can view its parameters and training materials.
The startup was co-founded by Armen Aghajanyan and Akshat Shrivastava, who previously worked at Fundamental AI Research (FAIR), Meta’s AI research arm. The duo sees their software as the future of industrial automation.
“Today’s physical AI forces a false choice: a generalist foundational model that requires multiple dedicated cloud GPUs per instance, or a narrow model that handles perception or control, but never both,” the company said.
Aghajanyan and Shrivastava say their tool differs from existing models in the field because it is general-purpose, meaning it is not built for specific repetitive tasks. Instead, they say, models are designed to be flexible depending on the particular environment (or situation) in which they are placed.
In an interview, Shrivastava asked me to think about what goes into a simple physical process like organizing boxes. “Imagine a robot is now being deployed to sort packages. What tasks does the robot need to perform?”
Such a relatively simple task actually consists of many steps. The robot must first read the label on the package, perform spatial analysis to understand where the boxes are, and decide which one to pick up. If you are picking up a series of boxes, you need to plan which boxes to pick up and in what order.
Perceptron’s software is designed to help the robot find its way through each step of the process. To be clear, there is already software in the industry that can help machines perform most of these tasks, but there are few programs designed to do it flexibly.
Where does this algorithmic alchemy data come from?
Models like Isaac 0.5 learn operational skills by ingesting vast amounts of video training data. Perceptron says its new model was fed 1 million hours of so-called common videos to teach the algorithm to identify specific settings, visuals, and scenarios. The company also relied heavily on so-called ego videos (videos shot from the perspective of a person completing a physical task, usually via a GoPro or wearable camera) and UMI videos, which are similarly used to teach AI systems movement by recording repetitive human movements.
Perceptron did not disclose the source of its training data, but Shrivastava said the company has “built a petabyte-scale dataset internally across modalities across the robot’s trajectory, including images, text, and video.”
The utility of software that can help robots operate competently within warehouses is clearly vast, and Perceptron believes it is well-positioned to lead the automation wave. The startup is poised to sell its software to a variety of vendors, potentially allowing the intelligence layer to be integrated into a wide range of industries.
These industries include manufacturing, logistics, warehousing, security, mobility, and even media and entertainment.
“Nothing like this really exists,” Aghajanyan said. “We’re really excited.”
The company previously raised $16 million in 2024 from Bessemer Venture Partners, The Explorer Fund, and SmartGateVC, according to Pitchbook. TechCrunch understands that the startup is in the process of completing additional rounds.
A previous version of this article incorrectly reported recent funding details.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
