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.
More from Embodied
- AI-Era Consumer Electronics Startup Inanimate Releases Demo Day Video, Starts WORKS Launch Planning — genmon · 2026-09-24
- Ropedia brings HOMIE Gen2 humanoid demo and evaluation dataset to IROS 2026 — liuziwei7 · 2026-09-24
- Microsoft shows offloading robot inference to edge/cloud boosts success rates and battery life — Microsoft Research · 2026-09-24
- Apsara Conference shows embodied AI expanding beyond robots into everyday products — Shruti_0810 · 2026-09-23
- Tesla Optimus v3 renders reportedly leaked in Android app APK ahead of Q4 debut — CyberRobooo · 2026-09-23
- Meta to unveil first camera-free smart glasses today amid privacy backlash — Polymarket · 2026-09-23