AI is transforming nearly every aspect of our digital lives, but its physical-world impact is still in its infancy. A new wave of startups aims to change that by bringing advanced machine learning out of the cloud and onto the factory floor. Perceptron, co-founded by two former Meta research scientists, is one such company.
Perceptron: A New Vision for Industrial AI
Founded in November 2024, Perceptron develops frontier vision models designed to help machines understand and interact with their physical environments. This week, the company launched Isaac 0.5, its latest model, which its creators say enables machines to “perceive, reason, and act” in industrial settings. The software is designed to help vision-guided robots navigate complex environments like warehouses and factory floors, and it also extracts visual intelligence from video captured by these robots. Isaac 0.5 is released as an open-weight model, meaning its parameters and training materials are publicly accessible for inspection.
Funding and Founding Team
The startup recently raised $21 million in a funding round led by Bessemer Venture Partners. It was co-founded by Armen Aghajanyan and Akshat Shrivastava, both of whom previously worked at Meta’s Fundamental AI Research (FAIR) division. The founders view their software as the future of automated industrial deployment, addressing what they call a “false choice” in physical AI today: either generalist foundation models that require multiple dedicated cloud GPUs for each instance, or narrow models that handle either perception or control—but rarely both.
A Flexible Approach to Robot Brains
Aghajanyan and Shrivastava say Isaac 0.5 stands out because it is general-purpose, not built for one specific repetitive task. Instead, the model is designed to adapt flexibly to different environments and situations. In an interview, Shrivastava illustrated this by describing a simple process like sorting boxes: a robot must read labels, perform spatial analysis to understand where boxes are, decide which to pick up, and plan the order of operations. While existing software can handle many of these tasks individually, few programs are designed to do them all flexibly in an integrated fashion.
Training on a Million Hours of Video
To teach the model these operational skills, Perceptron fed Isaac 0.5 a million hours of general video, helping its algorithm recognize specific settings, visuals, and scenarios. The company also relied heavily on egocentric video—footage captured from a first-person perspective using wearable cameras like GoPros—which provides a natural view of tasks as performed by humans. This training approach enables the model to learn from real-world demonstrations, rather than relying solely on synthetic data or rule-based programming.
Looking Ahead
As we move further into 2026, the demand for flexible, efficient industrial AI continues to grow. Startups like Perceptron are betting that open-weight models capable of both perception and control will become essential tools for factories and warehouses striving to automate complex workflows. By offering a general-purpose solution that can be deployed without massive computational overhead, Perceptron aims to bridge the gap between digital intelligence and physical action—bringing AI closer to the real world.
via TechCrunch AI