Isaac 0.5 (Perceptron)

A 36-billion-parameter open-weight robot foundation model from Perceptron, released in 2026, unifying seeing, thinking, and acting.

Isaac 0.5 is a robot foundation model the US startup Perceptron released with open weights under the Apache 2.0 license in August 2026; the company was founded by a team that had worked on Meta's Chameleon multimodal model, and had previously released the 1B–2B-parameter perception-language models Isaac 0.1 and 0.2. It has no relation to NVIDIA's Isaac platform. The model has 36 billion parameters total, built on the Qwen family of vision-language models as its backbone with added sparse mixture-of-experts layers; it takes in images, video, instructions, robot state, and action history, and can output text, coordinates, task progress, or actions — continuous actions are generated by a flow-matching expert, and discrete actions use FAST action tokens. The company says its training data includes 100,000 hours of experience from more than 35 robot types plus 1 million hours of ordinary video.

ExampleThe company's own fitted data-mixing law: to reach the same action-prediction loss, with only 1,000 hours of ordinary video you need about 5,900 hours of teleoperation data, but once ordinary video reaches 1 million hours, only about 28 hours are needed.

Also called

Perceptron Isaac, Isaac 0.1, Isaac 0.2

Related

Vision-Language-Action Model · Embodied Foundation Model · Open-weight Model · Flow Matching · Scaling Law · π0-FAST

Sources

PerceptronAI/Isaac-0.5 模型卡 (Hugging Face) (Chinese)
PerceptronAI/Isaac-0.1 模型卡 (Hugging Face) (Chinese)

As of

2026-08