Two ex-Meta FAIR research scientists have quietly turned their ambient-AI startup into a $21 million-funded factory-floor bet. Perceptron, founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, disclosed a $21 million funding round led by Bessemer Venture Partners alongside the launch of its Isaac 0.5 open-weight vision model on August 26.
A vision model built for the plant, not the browser
Isaac 0.5, released as an open-weight model, is designed to help vision-guided robots and cameras "perceive, reason and act" in unstructured industrial environments—warehouses, factory floors, logistics hubs, security perimeters. The model was trained on more than a million hours of general video plus egocentric (ego) footage recorded from head-mounted GoPros and UMI trajectories that capture repetitive human hand motion.
Perceptron's pitch is that physical AI today forces a false choice: either massive foundation models that demand multiple cloud GPUs per instance, or narrow purpose-built nets that handle perception or control but never both. Isaac tries to bridge that gap with a single general-purpose model that stays flexible across environments and tasks.
Why factories are the wedge
In an interview, Shrivastava walked through a simple example: a robot sorting packages needs to read labels, spatially reason about box positions, plan pick order and execute grasps in sequence. Existing task-specific software can handle each step; almost none does the full sequence flexibly. Perceptron says its intelligence layer can drop into manufacturing, logistics and warehousing, security, mobility and even media and entertainment workflows.
Money and momentum
Bessemer led the newly disclosed $21 million, layering on top of an earlier $16 million round from Bessemer, The Explorer Fund and SmartGateVC in 2024. TechCrunch reports Perceptron is in the process of closing an additional round on top of that.
The wider physical-AI push
The launch lands in a market where physical-AI funding is smashing records. Robotics AI leader Generalist AI closed another $200 million at a $3 billion valuation, autonomous freight startup Gatik banked $200 million in a Series D, and Hugging Face and Pollen Robotics just started shipping an open-source robot for developers to fine-tune. The story running through all of them: an AI intelligence layer, not the robot body, is the durable moat.
Reporting based on coverage from TechCrunch, VentureBeat, MarkTechPost and Perceptron's own release notes.