GR00T 2's Data Pyramid: 400k hrs human video vs just 23.5k hrs robot teleop
yukez · x · 2026-11-10
NVIDIA explains GR00T 2's Data Pyramid training recipe: much of what the robot learns comes not from robot data but from 400k+ hours of egocentric human video, 10k hours of handheld UMI demos, and only 23.5k hours of robot teleoperation.
GR00T 2 is described as a new generation of world-action models: it imagines how the physical world will unfold and generates the actions to get there, with one model running on parallel grippers, dexterous hands, and humanoids.
Related event: NVIDIA Open-Sources GR00T 2, a World-Action Model for Robots(20 posts)→
More from Embodied
- Wired tries the OpenAI-powered Microduck, a tiny robot duck that chases cats — Wired AI · 2026-11-12
- RoboCap-10k released: 10,000 hours of open egocentric robot data with metric actions — databoydg · 2026-11-12
- Claude Fable 5.2 Solves Rubik's Cube on Trossen WidowX Robot Arms in Anthropic Launch Demo — kscottz · 2026-11-12
- Scoble: AI glasses will get smaller and cheaper, and China's market appetite dwarfs the US — Scobleizer · 2026-11-12
- Blackmagic launches 108-mic 7th-order ambisonic microphone priced at $1,909 — Scobleizer · 2026-11-12
- NVIDIA unveils GR00T 2 humanoid open world action model, benchmarked on AI2's MolmoSpaces — Kyle_L_Wiggers · 2026-11-12