Most industrial robots today are, in a real sense, blind. They follow scripts. Perceptron thinks that’s the core problem worth solving, and it has $21 million and two ex-Meta scientists to back that bet.
According to TechCrunch, Perceptron this week released Isaac 0.5, a vision model built specifically for industrial environments like warehouses and factory floors. The company was founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both veterans of Meta’s Fundamental AI Research division. Bessemer Venture Partners led the seed round.
The pitch is straightforward. Existing industrial AI forces a tradeoff: either you use a large foundation model that demands serious cloud compute for every deployment, or you use a narrow model trained on one specific task. Perceptron says Isaac 0.5 does neither. It’s designed to be general-purpose, meaning it can adapt to different environments and task sequences rather than repeating a single fixed behavior.
To illustrate why that matters, co-founder Shrivastava walked through something as basic as sorting packages. A robot handling that job needs to read a label, understand the spatial layout of the boxes around it, decide which one to grab, and then plan the full sequence. That’s not one task. It’s four or five, chained together. Most current systems handle parts of this. Few handle all of it flexibly, without being reprogrammed for each variation.
The model was trained on a million hours of video, spanning general footage, ego-perspective video shot from wearable cameras, and UMI video that captures repetitive human physical actions. Shrivastava says the company has built petabyte-scale internal datasets covering images, text, video, and robotic trajectories. Training sources are not being disclosed publicly.
Isaac 0.5 is also being released as an open-weight model, which means researchers and developers can inspect both its parameters and its training methodology. That’s a meaningful move in a space where most industrial AI vendors treat their models as black boxes. It lowers the barrier for evaluation and builds credibility with technical buyers.
The broader context here is worth taking seriously. Physical AI has been a talking point for years, but practical deployment in unstructured environments has remained limited. Competitors like Covariant, Physical Intelligence, and Figure AI are all approaching the same problem from different angles. What separates Perceptron’s framing is the focus on visual intelligence as the primary layer, rather than end-to-end robotic control. The target markets include manufacturing, logistics, security, mobility, and media production.
Whether Isaac 0.5 performs as described in real deployments is still an open question. But the founding team, the funding, and the open-weight release together make this worth watching closely.
