Perceptron unveils Isaac 0.5 for factory robots
The startup founded by two former Meta FAIR researchers says its open-weight model is built to help robots perceive, reason, and act in warehouses and factories.
Perceptron, a startup founded by two former Meta research scientists, has launched a new AI model called Isaac 0.5 for robots working in industrial settings. The company says the system can help machines perceive, reason, and act in places such as warehouses and factory floors. It is also releasing the model as open-weight, meaning its parameters and training materials can be inspected by others.
The launch matters because it reflects a broader push in AI beyond software on screens and into physical environments. Perceptron is betting that visual AI can become a core layer in industrial automation, especially where robots need to do more than one narrowly defined job. That also raises practical questions about how adaptable these systems really are, what data they were trained on, and how easily they can fit into existing workflows.
Isaac 0.5 targets robots in changing industrial spaces
Perceptron was founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both former researchers at Meta’s Fundamental AI Research, or FAIR, division. The company says it is focused on frontier vision models built to help machines interact more effectively with the world around them.
According to Perceptron, Isaac 0.5 is aimed at vision-guided robots operating in complex environments. In practical terms, that means software meant to support robots moving through warehouse aisles, factory floors, and similar spaces where objects, obstacles, and tasks can shift from one moment to the next.
The startup presents the model as part of its vision for industrial automation. Its founders say the software is designed not only to recognize what it sees, but also to help machines interpret a situation and act on it.
Perceptron says the model links perception with action
Perceptron says Isaac 0.5 is designed to support several stages of a physical task. A robot sorting boxes, for instance, may need to read labels, understand where items are located, decide what to pick up, and plan the sequence of actions. The company says its model is intended to help robots move through that chain.
The startup argues this is where current systems still fall short. It says some existing models are strong at perception, while others handle control better, but few are built to do both flexibly. Perceptron also says many generalist foundation models need multiple dedicated cloud GPUs for each instance, which it frames as a limitation for physical AI deployments.
Aghajanyan and Shrivastava say their approach differs because it is general-purpose rather than built for a single repetitive task. In their account, the model is meant to adapt to the environment where it is deployed instead of being locked into one use case.
Perceptron keeps the training data sources private
Perceptron says Isaac 0.5 was trained on large volumes of video data. The company says it used a million hours of general video to teach the model to identify environments, visuals, and scenarios.
It also relied on ego video — footage captured from a person’s point of view, often through a GoPro or wearable camera, while completing a physical task. The company also used UMI video, another type of recording used to teach AI systems movement by showing repetitive human actions.
Perceptron is not disclosing the sources of that training data. Shrivastava said the company had “internally built petabyte-scale datasets that span across modalities, whether it’s images, text, video, etc. all the way through robotic trajectories.”
As IT-PUB News notes, that scale makes the project sound ambitious, but it also leaves some important details unanswered. Perceptron is making a large claim about the breadth of its data while keeping the underlying sources private.
Open-weight release broadens scrutiny and potential use
One of the most notable parts of the announcement is Isaac 0.5’s open-weight release. In simple terms, that means outside researchers and companies can inspect the model’s parameters and training materials. That could make the system easier to study and, potentially, easier to evaluate.
Perceptron recently raised $21 million in a funding round led by Bessemer Venture Partners. The startup says it wants to sell its software to multiple vendors, with the goal of integrating its intelligence layer across a range of industries.
Those industries, the company says, include manufacturing, logistics and warehousing, security, mobility, as well as media and entertainment. That suggests Perceptron sees the technology as more than a warehouse tool, even if industrial robotics is the clearest near-term use case in this launch.
If Isaac 0.5 performs as Perceptron says it can, it could help robots take on more complex physical work without such rigid programming. At the same time, the model’s general-purpose design, large-scale training setup, and open-weight release are likely to draw close attention from companies weighing whether tools like this are ready for real-world deployment.
