Perceptron, a physical-AI startup founded by two former Meta research scientists, launched Isaac 0.5 this week and disclosed a $21 million funding round led by Bessemer Venture Partners. The open-weight vision model is aimed at industrial settings — warehouses, factory floors, logistics hubs — where robots have to perceive an unstructured environment, reason about it, and act. Perceptron is releasing the model's parameters and training materials openly, a pointed contrast to the closed frontier labs that have dominated 2026's robotics-brain race.
The company was started in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both formerly at Meta's Fundamental AI Research division, known as FAIR. Isaac 0.5 was trained on 1 million hours of general video, supplemented by ego video captured through GoPro-style wearable cameras and UMI video of repetitive human actions. Together with what Shrivastava describes as petabyte-scale internal datasets, that mixture is meant to teach the model both what industrial environments look like and how humans move through them.
The pitch is that existing physical-AI stacks force a tradeoff between capability and cost. Perceptron argues its model does neither.
“Physical AI today forces a false choice: generalist foundation models that need multiple dedicated cloud GPUs for every instance, or narrow models that handle perception or control, but never both.”— Perceptron, Company statement
Key facts
- 01Perceptron raised $21 million in a round led by Bessemer Venture Partners and released Isaac 0.5 as an open-weight vision model.
- 02Co-founders Armen Aghajanyan and Akshat Shrivastava previously worked at Meta's Fundamental AI Research division and started Perceptron in November 2024.
- 03Isaac 0.5 was trained on 1 million hours of general video plus ego video from GoPro-style wearables and UMI video of repetitive human tasks.
- 04The company built petabyte-scale internal datasets spanning images, text, video, and robotic trajectories.
- 05Perceptron is targeting manufacturing, logistics, warehousing, security, mobility, and media as launch verticals.
The company frames Isaac 0.5 as general-purpose rather than task-specific — trained to adapt to a new environment rather than to run one repetitive motion. Shrivastava uses package sorting as the illustrative case: a robot has to read the label, do spatial analysis to locate boxes, decide which one to grab, and plan a sequence when there are several. Narrow perception or control models can handle individual steps. Fewer can string them together in a new warehouse without weeks of tuning.
"Imagine there's a robot being deployed to sort packages right now. What are the tasks it would need to do?" Shrivastava said in an interview. The list — read, localize, decide, sequence — is the kind of chain that has kept factory automation heavily scripted and brittle for two decades.
The data story is where Perceptron is doing the most differentiated work. Ego video and UMI video are useful precisely because they capture human movement from a first-person perspective, which maps more naturally onto how a robot arm or mobile platform experiences a scene than third-person surveillance footage does. Perceptron isn't disclosing where the 1 million hours came from.
Bessemer led the $21 million round. That is a small check by 2026 physical-AI standards — Generalist just closed a $200 million extension at a $3 billion valuation, which we covered earlier this month — but the open-weight release changes the distribution math. Any robotics integrator, manufacturer, or logistics operator can inspect Isaac 0.5, fine-tune it on their own workflows, and deploy it without a per-instance cloud GPU bill.
Perceptron says it is ready to sell into manufacturing, logistics and warehousing, security, mobility, and media and entertainment. That is a wide surface area for a company that just came out of stealth, and it depends on Isaac 0.5 doing what the founders claim: transferring across environments without task-specific retraining.
The skeptical read is that open-weight vision models for robotics have been announced before and have generally underperformed closed systems on real-world reliability. A million hours of video is a lot, but it is a fraction of what the largest robotics labs have accumulated, and video alone doesn't capture the physics of contact, friction, and failure. Perceptron will need to show benchmark results and customer deployments before the industrial buyers it is targeting — who care about uptime, not novelty — take the model into production.
The interesting bet Perceptron is making is that the physical-AI market will bifurcate the same way the language model market did: a handful of closed frontier systems from well-funded labs, and a growing open-weight tier that eats the long tail of practical deployments. If Isaac 0.5 is good enough to run a package-sorting cell or a pallet-scanning workflow without a cloud round-trip, the economics tilt hard toward whoever is willing to give the weights away. That is a much harder pitch for the $3-billion-valuation robotics labs to answer than it is for a $21 million seed-stage team with nothing to protect.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
