Pomegra
Bessemer leads a $21M seed into Perceptron AI as the ex-Meta FAIR startup ships Isaac 0.5, an open-weight model guiding robots through unstructured environments.
- Perceptron AI, founded November 2024 by ex-Meta FAIR scientists, raised $21 million in a seed round led by Bessemer Venture Partners.
- Isaac 0.5 is a 36B-parameter dynamic MoE model trained on 3 trillion tokens and over 100,000 hours of robot experience across 35+ systems.
- The open-weight model tops all five benchmark families it enters, at up to 8.5x lower inference cost than the top open comparator.
Lead
Perceptron AI, a Bellevue, Washington startup founded in November 2024 by two former Meta Fundamental AI Research scientists, disclosed a $21 million seed round led by Bessemer Venture Partners on August 26, 2026. The announcement arrived alongside the release of Isaac 0.5, a 36-billion-parameter open-weight embodied foundation model targeting warehouses and factory floors - environments where the robots guiding goods and assembling parts still struggle to see the unstructured world reliably.
What Does Isaac 0.5 Actually Do?
Isaac 0.5 is not a chatbot wrapper with a camera. The model fuses three historically separate capabilities - video understanding, embodied reasoning, and robot control - into a single sparse backbone trained on three trillion multimodal tokens. That training corpus included one million hours of general video, one million hours of ego-camera footage from GoPro-style wearables, and 100,000 hours of robot experience drawn from more than 35 distinct robot platforms. The result is a system that enables a robot arm or mobile logistics unit to watch its environment, interpret what it sees, and act without requiring a separate pipeline for each stage.
Perceptron's benchmark data is notable. Isaac 0.5 leads across all five benchmark families the company tested, at up to 8.5 times lower inference cost than the strongest available open comparator. Perhaps more striking for industrial buyers: the model cut the number of teleoperation demonstrations needed to train a well-calibrated action policy from 5,900 hours down to 28.
Why Physical AI, Why Now?
Warehouse and factory operators are under pressure to automate faster than skilled labor pools can replenish. Conventional machine vision systems fail when the environment changes - a misplaced box, a new SKU shape, a worker's hand unexpectedly in frame. Foundation models offer a path to general-purpose perception that doesn't require retraining every time the physical world shifts.
Physical AI has attracted billions in 2025 and 2026. One peer company, Generalist, closed a $200 million extension this year at a $3 billion valuation. Perceptron's $21 million check is small by comparison - a deliberate bet that an open-weight release can build adoption and a data flywheel before a larger round becomes necessary.
Co-founders Armen Aghajanyan (CEO) and Akshat Shrivastava (CTO) spent years at Meta's FAIR laboratory, the division responsible for foundational research behind models including Llama and DINO. Their research pedigree is central to Bessemer's thesis here: researchers who understand large-scale model training, pointed at the robot problem.
Open Weight as a Go-to-Market Decision
Releasing Isaac 0.5 under an open-weight license is a distribution choice as much as a technical one. The physical AI market lacks a text-model equivalent of the open-source community on Hugging Face. Most deployed robot systems rely on proprietary sensor stacks and vendor-locked software. An open-weight model invites industrial integrators, robotics hardware vendors, and university labs to evaluate the system without a sales cycle, building the usage data Perceptron needs to train subsequent versions.
The risk is speed. Proprietary competitors can iterate without managing an external community. The 8.5x inference cost advantage buys time, but the window is not permanent.
Target Markets and What It Displaces
Perceptron is targeting manufacturing, logistics and warehousing, security, mobility, and media and entertainment, with near-term focus squarely on the industrial stack. The near-term industrial target means competing directly against purpose-built machine vision vendors that have operated with narrow, deterministic models for years. Isaac 0.5's general capability should outperform those systems in edge cases - precisely where warehouse robots fail most visibly and most expensively.
Outlook
Perceptron now has $21 million and an open-weight model in the market. The real test arrives when early customers deploy Isaac 0.5 on hardware they didn't buy from Perceptron, in facilities that look nothing like a controlled benchmark. A follow-on round will likely depend on those deployment results. Bessemer's continued involvement across Perceptron's rounds signals conviction, but the physical AI market is moving fast enough that pedigree alone won't hold a position without production evidence behind it.
