← Leaderboard

GR00T 2

World-action
NVIDIA · Nov 9, 2026

World-action model that denoises future video and actions together with one 2B transformer, pretrained on about 400k hours of egocentric human video, 10k hours of UMI data and 23.5k hours of teleoperation; a lighter GR00T 2 Turbo runs on the G1’s Jetson Thor. Its whitepaper self-reports 67.2% on RoboCasa365, 35.5% on RoboDojo and 47.4% on RoboLab, none yet on the official boards; code is announced for December 2026.

Type
World-action
Parameters
2B
Backbone
Cosmos-Predict2.5-2B video diffusion transformer
Weights
Unknown
License
–
Commercial use
Unclear
Released
Nov 9, 2026
Robots
YAM, DROID Franka, Tianji + Sharpa hands (Dexter), Wuji hands, Unitree G1
SO-101
–
LeRobot
–
Inputs
RGB · P · L
VRAM
–

Benchmark results

3 of 13 benchmarks
RoboLab-120Simulation
Success by difficulty
Simple45.6%Moderate54.6%Complex37.6%
#5 of 2347.4%, self-reported by the authors, not a leaderboard entry
RoboCasa365Simulation
#2 of 2767.2%, self-reported by the authors, not a leaderboard entry
RoboDojo (sim)Simulation
#9 of 5435.5%, self-reported by the authors, not a leaderboard entry
Not reported on 10 benchmarks
Wrong or missing number? Report an error