Perceptron Unveils Isaac 0.5, a Visual AI for Factories
August 26, 2026
Perceptron, a startup founded by two former Meta research scientists, has announced Isaac 0.5, a visual AI model built for factories and warehouses.
The model aims to give machines the ability to “perceive, reason and act” within a single architecture, and it is slated for release as an open-weight model.
Here is what the model does and what it was trained on.
One Model for Perception, Reasoning, and Action
As TechCrunch reported on August 26, Isaac 0.5 is designed to help vision-guided robots navigate complex environments like warehouses and factory floors, and to extract visual intelligence from the video those robots record.
According to Perceptron’s official model card, Isaac 0.5 is a 36-billion-parameter sparse model that integrates video understanding, embodied reasoning, spatial grounding, task-progress estimation, and robot control in a single architecture.
Its defining trait is that it is general-purpose rather than built for one repetitive task. Speaking to TechCrunch, co-founder Akshat Shrivastava points to package sorting: a robot must read labels, map where boxes sit, and plan its picking order — a sequence few programs handle flexibly.
From Ego Video to Robot Experience: What Isaac 0.5 Learned From
Isaac 0.5 was trained on 1,000,000 hours of general video, plus ego video — first-person footage from wearable cameras — and UMI (Universal Manipulation Interface) data — human demonstrations of object manipulation, captured with handheld gripper devices and used to train robot policies.
The model card says training drew on 100,000 hours of robot experience across more than 35 robot systems, totaling three trillion multimodal tokens. Perceptron is not disclosing data sources, though Shrivastava says the company built petabyte-scale datasets internally.
Two Ex-Meta Scientists Take On a False Choice
Perceptron was founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both formerly of Meta’s Fundamental AI Research (FAIR) division. They join a wave of researchers leaving big labs to build startups — a trend that continued into 2026 with Jeff Dean’s departure from Google to launch Discovery Loop.
Perceptron says current physical-AI systems tend to fall into two camps: compute-heavy general models and specialized systems that cover only one side of the perception-control stack. Isaac 0.5 is designed to bridge that gap.
According to TechCrunch, the startup raised $16 million from Bessemer Venture Partners and others in 2024, per Pitchbook, and is reportedly closing an additional round. Target industries span manufacturing, logistics and warehousing, security, mobility, and media.
What to Watch Next
The model card lists the weights as coming soon, so the timing of the full open-weight release is the immediate question. The additional funding round TechCrunch reported also remains unconfirmed in size and backers.
Adoption is the longer-term test: Perceptron has not publicly named any customers or vendor integrations for Isaac 0.5 yet.
Sources
Perceptron – Isaac 0.5 Model Card (Hugging Face)
TechCrunch – Ex-Meta scientists want to bring visual AI to the factory floor
