Isaac 0.5: open-source 36B embodied foundation model trained on 1M hours of video, weights released

AkshatS07 · x · 2026-09-03

Perceptron released Isaac 0.5, an open-source embodied foundation model it calls the first open model at the frontier of multimodal video understanding, embodied reasoning, and robot control. Key facts: a 36B-parameter sparse model ingesting images, video, language, robot state and past actions; trained on 100K hours of robot experience, 1M hours of general video, 3T multimodal tokens across 35+ embodiments. They propose a scaling law for video-vs-teleop data mixes: scaling general video from 1K to 1M hours cut teleoperation needed to hit the same held-out action loss from 5,900 hours to 28. Weights and paper are public; YAM and SO-101 checkpoints are coming.

Original post →

More from Embodied

Embodied channel →