Perceptron is trying to move AI from screens into industrial spaces where machines must understand what is around them and decide what to do next. The startup, founded in November 2024 by former Meta research scientists Armen Aghajanyan and Akshat Shrivastava, has introduced Isaac 0.5, a vision model built for settings such as warehouses and factory floors.
The company says Isaac 0.5 is designed to help machines perceive, reason and act in industrial environments. That makes the model part of a broader push to bring visual AI into physical operations, where robots must deal with objects, layouts, movement and changing surroundings.
A model built for industrial robots
Isaac 0.5 is aimed at vision-guided robots that need to move through complex environments. In practical terms, that can mean helping a robot navigate a warehouse, interpret what it sees on a factory floor, or turn video captured by robots into useful visual intelligence for companies.
The company is also releasing Isaac 0.5 as an open-weight model. That means its parameters and training materials can be inspected by anyone, a notable choice for a company working on software intended for physical deployment.
Perceptron is not presenting Isaac 0.5 as a single-purpose tool for one repetitive job. Aghajanyan and Shrivastava say the model is meant to be general-purpose, with flexibility across different environments and situations. That distinction matters because the company argues that many current options force a tradeoff between large generalist systems and narrower models focused on only part of the task.
The company describes that tradeoff directly: physical AI, in its view, often asks users to choose between foundation models that require multiple dedicated cloud GPUs for every instance and specialized models that handle perception or control, but not both.
Why simple warehouse work is not simple for a robot
Shrivastava used package sorting to explain the challenge. A person may see organizing boxes as a routine task, but a robot must break it into a sequence of perception, spatial reasoning and planning decisions.
For a robot sorting packages, the work can include several steps:
- Reading the label on a package.
- Understanding where boxes are in space.
- Choosing which box to pick up.
- Planning the order of pickup when there are multiple boxes.
Each step depends on the robot understanding visual information from the physical world. A system that only recognizes objects is not enough by itself. A system that can plan motion but lacks strong perception also leaves gaps.
Perceptron says its software is designed to help robots move through those linked steps. The source article notes that software already exists for many of these individual tasks, but fewer programs are designed to handle them flexibly across changing situations.
The data behind Isaac 0.5
Models like Isaac 0.5 learn from large quantities of video training data. Perceptron says the new model was trained on a million hours of general video, which was used to help the algorithm recognize settings, visuals and scenarios.
The startup also relied heavily on ego video. In the source article, ego video is described as video captured from the perspective of a person performing a physical task, often through a GoPro or wearable camera. That kind of footage can show the visual flow of real-world work from the position of the person doing it.
Another source of training material was UMI video. These videos are used to teach AI systems movement by recording repetitive human actions. For a system intended to support robots, that type of data is relevant because industrial tasks often involve repeated physical motions that still require perception and timing.
Perceptron has not disclosed the sources of its training data. Shrivastava said the company had internally built petabyte-scale data sets spanning images, text, video and robotic trajectories.
Where Perceptron wants the technology to go
Perceptron recently raised $21 million in a funding round led by Bessemer Venture Partners. The company is now preparing to market its software to a variety of vendors, with the aim of having its intelligence layer integrated into multiple industries.
The industries named by the company include manufacturing, logistics and warehousing, security, mobility, and media and entertainment. Those categories all involve visual information, physical environments, or both. They also suggest that Perceptron sees Isaac 0.5 not only as warehouse robot software, but as a broader visual AI layer for machines and systems operating outside purely digital settings.
The founders bring experience from Meta’s Fundamental AI Research (FAIR), the tech giant’s AI research division. With Isaac 0.5, they are applying that research background to industrial automated deployment, where the model must be useful in places that are messy, spatial and operational rather than only informational.
Aghajanyan summed up the company’s view by saying that nothing like this really exists out there. Whether Isaac 0.5 becomes a widely used layer for physical AI will depend on how well it performs in the environments Perceptron is targeting. For now, the launch shows how quickly visual AI is moving toward robots, warehouses and factory floors.