Perceptron, Inc.
"A layer of intelligence for the physical world."
Bellevue foundation-model lab founded by two ex-Meta FAIR scientists; its open-weight Isaac models give robots video understanding, reasoning and control.
Perceptron AI is a foundation-model research company building what it calls Physical AGI: perception and reasoning models intended to let machines understand and act in the physical world rather than only in text and static images. It was founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava, both former research scientists at Meta's Fundamental AI Research (FAIR) lab, where they worked on multimodal model research. Aghajanyan is chief executive and Shrivastava chief technology officer. The company is based in Bellevue, Washington, and operates as a small research-and-engineering team rather than a hardware manufacturer; it supplies models, APIs and licensed weights to the companies that build and operate machines.
The company's work centres on the Isaac family of open-weight models and the closed Mk1 vision-language model. Isaac began as a compact perceptive-language model, with Isaac 0.1 at roughly 3B parameters and Isaac 0.2 preview builds at 1B and 2B, aimed at grounded visual question answering and object localisation at sizes small enough to run close to the machine. Isaac 0.5, released on 28 August 2026, is a considerably larger step: a 36-billion-parameter sparse mixture-of-experts model that Perceptron describes as an embodied foundation model, combining multimodal video understanding, embodied reasoning and robot control in a single backbone. Perceptron states it was trained on about three trillion multimodal tokens, including roughly one million hours of general video and 100,000 hours of robot experience drawn from more than 35 robot systems, with egocentric and Universal Manipulation Interface (UMI) data in the mix. The company reports a 97.2 per cent success rate on the LIBERO manipulation benchmark and positions the model against Physical Intelligence's pi-0.5 and NVIDIA's GR00T N1.7. Isaac 0.5 weights are published on Hugging Face and the fine-tuning and inference code is on GitHub under Apache 2.0, alongside a technical report.
Commercially, Perceptron sells access rather than machines. It offers hosted API access through platform.perceptron.inc, a separate Egocentric API for first-person video, and commercial licensing of model weights for customers that need to run the models on their own infrastructure. The company names manufacturing, logistics and warehousing, security, mobility, and media and entertainment as the sectors it serves; its own site lists a wider set of target domains including sports, infrastructure, agriculture, aerial platforms, wearables and smart home. No customer has been publicly named.
Perceptron raised 21 million US dollars from a group led by Bessemer Venture Partners with Foundation Capital, S32 and SmartGateVC participating. The financing was reported on 26 August 2026 alongside the Isaac 0.5 launch, and the company told press at the time that it was in the process of closing an additional round. Round nomenclature is inconsistent across sources: Crunchbase records the financing as a pre-seed round while press coverage has described it as seed or Series A, so the amount and the investor list are the dependable part of the record.
In a robotics supply chain Perceptron sits in the perception-and-control software layer rather than the machine layer. It belongs to a small group of companies, alongside Physical Intelligence, Skild AI and NVIDIA's GR00T programme, attempting to supply a general vision-language-action model that a robot builder can adopt instead of training task-specific perception from scratch. Its open-weight strategy is the main point of difference: publishing an embodied model at this scale under a permissive licence is uncommon in this segment and makes the models directly testable by integrators and system builders before any commercial commitment. The company is early stage, founded under two years ago, with a headcount in the single digits to low tens, no disclosed revenue and no named deployments, so its published models rather than its commercial footprint are the reason to track it.
Primary type & automation activities this supplier delivers:
AutoGauge
Inline non-contact automated gauge system for continuous dimensional inspection of automotive body-in-white and assembly line components.
More related news
- 1
- 2
- 3
- 4
- 5
-
6
IROS 2026 Category Award Finalists: All 22 Papers RobotToday · Oct 07, 2026
-
7
IROS 2026 by the Numbers: Papers, Countries, Topics RobotToday · Oct 07, 2026
-
8
IROS 2026 Best Paper Finalists: South Korea Leads RobotToday · Oct 07, 2026
- 9
- 10
- 11
-
12
RSS 2027 Athens: Jul 6–11 | Robotics: Science and Systems RobotToday · Oct 06, 2026
-
13
IROS 2027 Florence: Sep 26–Oct 1 | Intelligent Robots RobotToday · Oct 06, 2026
- 14
- 15
-
16
CICEE 2027: May 12-15, Changsha, China RobotToday · Oct 05, 2026
-
17
WRC 2026 Core Components: 137 Robot Parts Suppliers RobotToday · Aug 28, 2026
-
18
Vision-Based Tactile Sensing: The Race to Build What Tesla's Optimus Needs RobotToday · Apr 02, 2026
-
19
Space Robot- Building Skyscrapers in Space: ISMA's Industrial Revolution RobotToday · Feb 18, 2026
-
20
Flexible Machining of Large Structures (Part III) RobotToday · Jan 19, 2026
Contact Perceptron, Inc.
WEBSITE
https://www.perceptron.incPHONE
+1 734-414-6100HEADQUARTERS
United States
Company Facts
Founded
2024
Primary Role
Software/Algorithm
Company Size
employees 1-10
Primary Region
North America
Annual Sales
-
Funding Stage
Early-Stage VC
Funding Total
$ 21,000,000
