Perceptron's embodied foundation model: 10x more video pretraining substitutes 10x less teleop data

AI Engineer · youtube · 2026-09-24

In an AI Engineer talk, Perceptron AI's Armen Aghajanyan detailed their embodied foundation model: a perceptive objective that learns which visual tokens matter (only 2% of a million tokens carry loss), data-sparse mixture-of-experts routing to fight context bloat, detection as an agentic task, and a petabyte-scale training mix. Headline result: a new scaling law where 10x more video pretraining substitutes 10x less teleop data ($100/hr). Open weights promised for July.

Original post →

More from Embodied

Embodied channel →