Perceptron

Perceptron AI develops multimodal foundation models for physical AI — AI that perceives, reasons and acts in real-world environments. The company was founded in November 2024 by researchers from Meta's FAIR and is based in Bellevue, Washington, with a small team of AI researchers and engineers. It is backed by Bessemer Venture Partners.

Its flagship commercial model, Perceptron Mk1, is a vision-language model built for native video understanding and embodied reasoning, taking video at up to 2 FPS with a 32K-token context and returning structured time codes and clips. Alongside it the company develops the Isaac series of open-weight embodied foundation models, including Isaac 0.1, an edge-oriented Isaac 0.2 and the 36-billion-parameter sparse mixture-of-experts Isaac 0.5, aimed at robot control and embodied reasoning; the models support both a diffusion transformer for continuous control and a FAST action vocabulary for discrete control.

On top of the models Perceptron offers several APIs. Egocentric turns raw robot and egocentric video into structured, policy-trainable supervision with per-frame hand boxes, 21-keypoint skeletons, persistent left/right identity and per-hand action captions against a 16-action taxonomy. Agentic Detection grounds anything described in natural language or with a single example image, reasoning over an image more than once to resolve small, dense and ambiguous details, and generalizes across satellite, drone, robot, vehicle and wearable viewpoints. A Files API and a video_frames input type for Mk1 cut the overhead of handling large video payloads.

Developers and enterprises reach the models through the Perceptron Platform and API, a Python SDK, GitHub repositories and external model marketplaces, with Mk1 priced at $0.15 per million input tokens and $1.50 per million output tokens — the company says this matches or beats OpenAI, Anthropic and Google at 80–90% lower cost. Its systems are deployed with enterprise customers in manufacturing, media, logistics and security, and Isaac is released under Apache 2.0 with integration into the LeRobot robotics library.

The team also publishes research, including a study on formalizing visual reasoning that examined backtracking and spatial grounding in visual games, and work reporting a scaling law that reduces real-robot teleoperation data needs by roughly a factor of 210. It reports 88.5 on VSI-Bench for Mk1 and an average 97.2% success rate on LIBERO for Isaac 0.5.