Ex-Meta scientists' Perceptron brings vision AI to the factory floor

AI has largely remained confined to the digital realm, but a growing number of startups are looking to take it into the physical world. Perceptron, started by two former Meta research scientists, is one such company; founded in November 2024, it develops frontier vision models meant to help machines interact more competently with their physical environments.

This week the company launched its latest model, Isaac 0.5, which its creators say is designed to give machines the ability to perceive, reason, and act in industrial settings — specifically, helping vision-guided robots navigate complex environments like warehouses or factory floors, and helping companies extract visual intelligence from video recorded by those robots. Isaac 0.5 is being released as an open-weight model, so its parameters and training materials can be inspected by anyone.

Perceptron was co-founded by Armen Aghajanyan and Akshat Shrivastava, who previously worked at Meta’s Fundamental AI Research (FAIR) division. The pair see their software as the future of automated industrial deployment: “Physical AI today forces a false choice: generalist foundation models that need multiple dedicated cloud GPUs for every instance, or narrow models that handle perception or control, but never both,” the company says. Aghajanyan and Shrivastava argue their tool differs from existing models because it is general-purpose rather than built for one specific, repetitive task, and is designed to flex depending on the environment it’s operating in — Shrivastava gave the example of a robot sorting packages, which must read a label, do spatial analysis to locate boxes, decide which to pick up, and plan the order if picking up several, steps that existing software can mostly already do individually but rarely flexibly, all together.

Isaac 0.5 learned its operational skills from a million hours of general video, plus what’s known as “ego video” — footage captured from the perspective of a person completing a physical task, typically via a GoPro or wearable camera — and “UMI video,” which similarly teaches AI systems movements by recording repetitive human actions. Perceptron isn’t disclosing its training-data sources, but Shrivastava said the company has “internally built petabyte-scale datasets that span across modalities, whether it’s images, text, video, etc. all the way through robotic trajectories.” The startup plans to market the software to vendors across manufacturing, logistics and warehousing, security, mobility, and media and entertainment. Perceptron previously raised $21 million from Bessemer Venture Partners, Foundation Capital, and S32, with SmartGateVC also participating in its founding round, according to Pitchbook; TechCrunch understands the startup is in the process of closing an additional round.

Why it matters

An open-weight, general-purpose vision model for robots — rather than task-specific software — is a bet that the bottleneck in warehouse and factory automation is flexibility across steps, not raw perception or control capability alone.