Vibeleaderboard
Perceptron

ai lab

Perceptron

Perceptron matters because it is trying to make one model representation useful across video understanding, spatial grounding, task progress, and robot action. Isaac 0.5 also gives the physical-AI field an open artifact and a concrete data-scaling hypothesis to inspect, instead of limiting the company to a hosted video demo.3,5,1

Models in the Atlas →

Profile

Overview

A FAIR team turns toward physical AI

Perceptron is a Bellevue, Washington physical-AI company founded in November 2024 by Armen Aghajanyan and Akshat Shrivastava. Both founders previously worked at Meta's Fundamental AI Research group, where Aghajanyan contributed to the Chameleon multimodal model program. Perceptron raised an initial $16 million and organized around a specific problem: models that can understand continuous visual scenes and eventually control machines inside them.1,2

From compact perception to commercial video

The company's first public Isaac releases were compact vision-language models for spatial grounding, pointing, optical character recognition, and visual question answering. Perceptron Mk1, launched as a hosted model in May 2026, shifted that work toward commercial video analysis. It accepts image and video input and can return text and time-coded answers. VentureBeat independently tested one public-domain film, but the article's benchmark and price comparisons otherwise came from Perceptron.6,2,7

Isaac joins perception and control

Isaac 0.5 expands the research program from perception into robot control. The open 36-billion-parameter sparse model reads video, language instructions, robot state, and prior actions, then produces language, coordinates, task-state estimates, or continuous and discrete actions. Its technical report describes joint training on one million hours of general video, 100,000 hours of robot experience, more than 35 robot systems, and three trillion multimodal tokens.3,4

A testable claim about video and robot data

The strongest research claim is a measured tradeoff between broad video and costly teleoperation data. Within Perceptron's training grid, increasing general video from 1,000 hours to one million reduced the teleoperation needed to reach a specified held-out action-loss threshold from 5,884 hours to 28. The company released weights, adaptation code, evaluation settings, and deployment configurations, but parts of the future-percept objective and data sources remain proprietary. Independent reporting confirms the launch and company history, not the full benchmark claims.3,5,1

Notable contributions

  1. 01A shared backbone for perception and controlIsaac 0.5 trains video understanding, spatial grounding, task-state estimation, and robot actions through one sparse representation. The contribution is this released system, not the broader idea of vision-language-action models.3,4
  2. 02Measured exchange between video and teleoperationThe Isaac report varies general, egocentric, handheld-gripper, and teleoperation data to estimate how abundant action-free video changes the amount of expensive robot demonstration needed at a fixed loss target.3
  3. 03Open multi-embodiment deployment materialsPerceptron paired the Isaac 0.5 weights with adaptation code, LeRobot integration, evaluation settings, and configurations for more than 35 robot embodiments represented in training.5,4