Nvidia unveils GR00T N2 world action model trained on 400k hours of human video

量子位 · wechat · 2026-11-10

At CoRL, Nvidia GEAR leads Jiaju “Kexin” Zhu and Jim Fan unveiled GR00T N2, a world action model that jointly predicts future video and robot actions, demoing tasks from cutting with scissors to installing GPUs and making breakfast. Model, data, simulation, and capture hardware will be open-sourced; 100k hours of human data are already available.

Key facts:

X-Adapt cross-embodiment transfer: human mocap demos alone gave Dexter 65% success on filter-paper placement; UMI handheld demos transferred to the YAM arm at 60% over 4-minute assemblies; DexUMI exoskeleton demos enabled breakfast prep
Sample-efficient: Dexter-50 zero-shot 40%, 71% with 1 demo, 95% with 5 demos (<3 min total), then 30 min LoRA fine-tuning on one RTX 5090
Pretraining mix: 402,730 hrs human egocentric video, 10,184 hrs UI demos (2,897 tasks), 23,500 hrs teleop data; ablations show human video lifts average task progress from 50% to 70%
Full toolchain: UMI/DexUMI capture, Human-1M + Flora data, SimFoundry simulation, ASPIRE/Turbo/SONIC execution (124ms per prediction on Jetson Thor, 50Hz whole-body control), plus a WAM API for reasoning agents to call physical skills

Jim Fan: “Astra didn't shatter our dreams — it shattered our walls. This isn't winter; it's robot summer.” Also noted: Claude Fable 5.2 scored 64.9% vs GPT Astra's 29.8% on Anthropic's RoboDojo sim dual-arm benchmark.

Original post →

Nvidia unveils GR00T N2 world action model trained on 400k hours of human video

More from Embodied