Two ex-Meta researchers want warehouse robots to finally understand what they're looking at
Perceptron's Isaac 0.5 is an open-weight vision model built to help industrial robots handle complex, multi-step tasks on their own. The startup raised $21 million and is targeting factories, warehouses and logistics hubs.
Key points
- Perceptron, founded in November 2024 by two former Meta AI researchers, launched its latest vision model, Isaac 0.5, this week.
- The model is open-weight, meaning anyone can inspect its internal settings and training materials.
- Perceptron raised $21 million in a funding round led by Bessemer Venture Partners.
- Isaac 0.5 was trained on one million hours of video to teach it how to understand real-world environments.
- Target markets include manufacturing, logistics, warehousing, security and media.
Sorting a box sounds simple. For a robot, it is not.
Before a warehouse machine can move a single package, it has to read the label, figure out where every box sits in three-dimensional space, decide which one to pick up first, and plan the whole sequence in order. Today, companies typically stitch together several separate software programs to cover each step. Perceptron, a startup founded in November 2024, says its new model does all of it in one.
What exactly did Perceptron release?
The company launched Isaac 0.5 this week: a vision model, meaning software that interprets what a camera sees and turns it into decisions a machine can act on. Unlike many existing tools built for one fixed job, Isaac 0.5 is designed to be general-purpose, adapting to different environments and tasks without being reprogrammed from scratch.
It runs on vision-guided robots, machines that use cameras rather than fixed sensors to find their way around. The software helps them move through warehouses or factory floors, and it can also pull useful information from video footage those robots have already recorded.
Isaac 0.5 is released as an open-weight model. That means the model's internal numerical settings, which encode everything it has learned, are publicly available for researchers and companies to inspect, test and build on.
Who built it, and who is paying for it?
Perceptron was co-founded by Armen Aghajanyan and Akshat Shrivastava, both formerly of Meta's Fundamental AI Research division, the team inside Meta dedicated to long-term AI science. The company recently closed a $21 million funding round led by Bessemer Venture Partners.
How did the model learn to understand the physical world?
Training a model like this requires enormous amounts of video. Perceptron fed Isaac 0.5 on one million hours of what the company calls general video to teach it to recognise different settings and objects. It also used ego video, footage shot from a person's point of view through a wearable camera like a GoPro, which shows the model how a human physically moves through a task. A third source was UMI video, recordings of repetitive human actions used to teach AI the mechanics of movement.
Perceptron has not disclosed where that footage came from. Co-founder Shrivastava said the company built internal datasets measured in petabytes (one petabyte is roughly a million gigabytes) covering images, text, video and recorded robot movements.
What does this mean for workers and businesses?
For companies running warehouses or factories, software like this could reduce the number of separate systems needed to keep a robot fleet running. For workers in those spaces, faster and more capable automation is already reshaping which jobs exist and which do not. This model's release does not change that picture overnight, but it points in a clear direction.
Perceptron says it plans to sell to vendors across manufacturing, logistics, security, mobility and media and entertainment, so the technology could surface in a wide range of products over the coming years.
Common questions
Is Isaac 0.5 a finished product companies can deploy right now?
Perceptron is actively marketing the model to vendors, but widespread deployment in real warehouses will take time as buyers test and integrate it into their existing systems.
What does open-weight mean for safety?
Open-weight models let outside researchers check for flaws or biases in the software, which many experts consider a plus. It also means anyone can adapt the model without Perceptron's approval, which raises separate questions about how it might be used.
