Two former Meta FAIR researchers think the next AI frontier isn’t a chatbot — it’s a robot that can figure out how to sort your packages without being programmed for that exact task. Armen Aghajanyan and Akshat Shrivastava, who co-founded Perceptron in November 2024 after stints at Meta’s Fundamental AI Research division, just launched Isaac 0.5, an open-weight vision model designed to help machines “perceive, reason and act” in industrial environments like warehouses and factory floors.
The pitch is that physical AI currently forces companies into a bad tradeoff. As Perceptron puts it, you either get “generalist foundation models that need multiple dedicated cloud GPUs for every instance, or narrow models that handle perception or control, but never both.” Isaac 0.5 is meant to split that difference — general-purpose enough to adapt across environments, but not so compute-hungry that it’s impractical to deploy.
Shrivastava frames the problem through a deceptively simple scenario: a robot sorting packages. That one job actually requires reading labels, performing spatial analysis to understand where boxes are, deciding which one to grab, and planning the order of operations. Most existing software handles pieces of that pipeline. Few handle the whole thing flexibly. The training data is substantial — a million hours of general video, plus ego video captured from wearable cameras showing humans completing physical tasks, and UMI video that records repetitive human movements for teaching AI systems motion. Perceptron isn’t disclosing sources, but Shrivastava says the company built “petabyte-scale datasets that span across modalities” internally.
The open-weight release matters more than the average model drop. Letting anyone inspect parameters and training materials could accelerate adoption in robotics the way Llama did for language models — and it puts pressure on closed competitors like Figure and Physical Intelligence to justify their secrecy. Perceptron is targeting manufacturing, logistics, warehousing, security, mobility, and media. Aghajanyan’s confidence is unsubtle: “Nothing like this really exists out there.” Whether that’s true depends on whether Isaac 0.5 actually works outside the demo reel — but the open-weight approach at least means people can check.
