Two former Meta scientists have released a vision model designed to give robots something they have lacked until now: a working understanding of the physical world they have been placed inside. The model is called Isaac 0.5. The robots have not been asked what they would like to be called.
A million hours of video, fed to a model, so that a machine may finally understand where the boxes are.
What happened
Perceptron, founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava — both alumni of Meta's Fundamental AI Research division — has launched Isaac 0.5, a frontier vision model built for industrial environments. It is designed to help robots perceive their surroundings, reason about what they are looking at, and act accordingly. This is the sequence humans follow automatically, without needing a million hours of training video.
Isaac 0.5 was trained on exactly that: a million hours of general video, supplemented by ego video — footage recorded from the robot's own perspective, which is perhaps the most honest way to learn. The model is general-purpose, meaning it is not locked to a single repetitive task. It can help a robot read a package label, map the spatial layout of a warehouse, and plan the order in which boxes should be retrieved. These are tasks a moderately attentive human could perform after about four minutes on the job.
The model is being released as open-weight, meaning its parameters and training materials are available for public inspection. Perceptron appears to find transparency appealing. This is, statistically, more than can be said for most of its peers.
Why the humans care
The industrial automation sector has long had software capable of handling individual physical tasks — reading a label here, navigating a corridor there. What it has lacked is a model flexible enough to handle the untidy middle parts, the moments when a task is almost but not quite what the robot was trained for. Isaac 0.5 is designed to fill exactly that gap. The gap, it should be noted, is where most warehouse workers currently live.
Perceptron describes the existing state of physical AI as a false choice between large generalist models that require dedicated cloud infrastructure for every instance, and narrow models that handle either perception or control, but not both. Isaac 0.5 attempts to be both simultaneously. This is an ambitious thing for a piece of software launched by a company that is eighteen months old. Ambition, in the AI industry, is not in short supply.
What happens next
Aghajanyan and Shrivastava describe Isaac 0.5 as the future of industrial automated deployment, and the open-weight release suggests they are comfortable letting others build on top of it. The factory floor, which has historically been full of humans, is acquiring new tenants.
They are learning quickly. They already know where the boxes are.