Perceptron is a startup founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both formerly researchers at Meta’s Fundamental AI Research (FAIR). The company develops frontier vision models intended to enable machines to interact more competently with physical environments.
This week Perceptron released its latest model, Isaac 0.5, which the company describes as designed to let machines “perceive, reason and act” in industrial settings. The software targets vision-guided robots operating in complex spaces such as warehouses or factory floors, and also helps extract visual intelligence from video recorded by those robots.
Isaac 0.5 is being published as an open-weight model, meaning its parameters and training materials can be inspected by outside parties.
Perceptron recently raised $21 million in a funding round led by Bessemer Venture Partners. The founders say their tool aims to address a perceived gap in physical AI: today’s options either require large cloud GPU resources for generalist foundation models or are narrowly focused on either perception or control. Perceptron positions Isaac 0.5 as a more general-purpose, flexible alternative.
Rather than being tuned for one repetitive task, the company says the model is designed to adapt to different environments and situations. Shrivastava illustrated the complexity with a simple example: a robot sent to sort packages must read labels, analyze spatial arrangements, decide which box to pick, and plan pick-up sequences — a series of distinct steps that Perceptron’s software aims to support.
Where the model’s knowledge comes from
Isaac 0.5 was trained on very large video datasets. Perceptron says it used roughly one million hours of so-called general video to teach the model to recognize settings, visuals and scenarios. The company also made heavy use of ego video — footage recorded from a person’s perspective with wearable cameras like GoPro — and UMI video, which captures repetitive human actions to teach movement patterns.
Perceptron has not disclosed detailed sources for its training data, 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.”
Commercial potential and target industries
Perceptron believes a general-purpose vision model that helps robots operate reliably in warehouses and factories has broad commercial value. The company plans to market its software to a range of vendors and expects the technology could be integrated across industries including manufacturing, logistics and warehousing, security, mobility, and media and entertainment.
“Nothing like this really exists out there,” said Armen Aghajanyan. The founders express enthusiasm about bringing their system into real-world industrial deployments.
Timing and practical implications
Founded in November 2024, Perceptron released Isaac 0.5 this week and has secured $21 million in funding to continue development. Publishing the model as open-weight allows researchers and industrial partners to examine the model and its training materials while the company pursues commercial collaborations to deploy the technology.
