Two former Meta research scientists just put a new stake in the ground for physical AI. Their startup, Perceptron, launched Isaac 0.5 this week, a frontier vision model built to give robots the ability to perceive, reason and act inside warehouses and factories, according to TechCrunch AI. As TechCrunch AI reports, the founders, Armen Aghajanyan and Akshat Shrivastava, both came out of Meta’s Fundamental AI Research (FAIR) division before starting the company in November 2024.
What stands out here is the target. Most of the AI wave has stayed inside screens and chat windows. Perceptron wants to move it into the real world, where a machine has to look at a messy shelf and actually do something about it.
What Isaac 0.5 does
The model is aimed squarely at industrial settings. Here’s what it brings to the table, based on the TechCrunch AI report:
- Navigation for vision-guided robots. It helps bots find their way through complex spaces like warehouse aisles and factory floors, not just one fixed route.
- Visual intelligence from bot footage. Companies can pull usable insights out of the video those robots record while they work.
- A general-purpose design. This is the key selling point. Instead of being locked to one repetitive job, the model adapts to the environment and situation it lands in.
- Open weights. Isaac 0.5 ships as an open-weight model, so anyone can inspect its parameters and training materials.
Shrivastava framed the pitch with a simple example: a robot sorting packages. It has to read the label, figure out where the boxes sit in space, decide which one to grab, then plan the order for the rest. Simple to describe, lots of steps to execute. Isaac is built to help a robot walk through each one.
Why the founders think it’s different
Perceptron argues that physical AI has been stuck with a bad trade-off. In the company’s words, teams face “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.”
Their answer is a model that’s flexible rather than single-task. To be fair, the industry already has software that handles most of these individual jobs. What’s rare, the founders say, is software that does them flexibly across changing conditions. “Nothing like this really exists out there,” Aghajanyan told TechCrunch AI.
Where the training data comes from
Models like this eat enormous amounts of video, and Isaac 0.5 is no exception. TechCrunch AI reports the model trained on roughly a million hours of general video to learn settings, visuals, and scenarios. The team leaned heavily on two specialized types:
- Ego video, footage shot from a person’s point of view with a GoPro or wearable camera while completing a physical task.
- UMI video, which teaches AI systems movement by recording repetitive human actions.
Perceptron isn’t naming its data sources, but Shrivastava said the company built “petabyte-scale datasets that span across modalities, whether it’s images, text, video, etc. all the way through robotic trajectories.”
Who it’s for and where it’s headed
The company plans to sell Isaac to a wide range of vendors, hoping its intelligence layer lands across several sectors. TechCrunch AI lists the targets: manufacturing, logistics and warehousing, security, mobility, and media and entertainment.
On funding, Perceptron previously raised $16 million from Bessemer Venture Partners, The Explorer Fund, and SmartGateVC in 2024, per Pitchbook. TechCrunch AI understands the startup is now closing an additional round.
One caveat worth keeping in mind: the undisclosed data sources. When a model is trained on a million hours of video and the origins stay private, buyers in regulated industries will have questions about provenance before they deploy it near expensive machinery or people.
This matters because the next real test for AI isn’t another chatbot. It’s whether these systems can operate reliably in a warehouse at 2 a.m. with nobody watching. Perceptron is betting a flexible, open-weight vision model is how you get there. Whether the robots agree is the part we’ll watch next. Full details are available at the original TechCrunch AI report.