NVIDIA Unveils GR00T 2, a Robot Model That Learns From Human Video

NVIDIA's GR00T 2 learns dexterous skills from a few demonstrations. Open weights are promised for December, and the benchmark numbers come from NVIDIA.

Published
· 2 min read
Category
AI
Topics
nvidiaroboticsgr00tisaac-lab
한국어 요약
NVIDIA가 11월 9일 로봇 학습 학회 CoRL에서 월드 액션 모델 GR00T 2와 로봇 시뮬레이션 학습 프레임워크 Isaac Lab 3.0을 공개했습니다. 사람의 1인칭 영상 10만 시간을 포함한 데이터로 학습해 케이블 꽂기 같은 정밀 작업을 배우는 것이 목표입니다. NVIDIA에 따르면 시연 없이 50개 과제 중 40개를 한 번 이상 성공했고, 과제당 시연 1회에 평균 71%, 5회에 95%의 성공률을 보였습니다. 다만 '한 번 이상 성공'은 느슨한 기준이며 수치는 모두 NVIDIA가 발표한 것입니다. '오픈'이라고 소개됐지만 가중치는 12월 공개 예정이고 라이선스는 아직 알려지지 않았습니다.

At the Conference on Robot Learning (CoRL) in Austin, Texas, NVIDIA introduced GR00T 2 on Nov. 9. It also released Isaac Lab 3.0, a new version of its open-source framework for training robots in simulation.

What GR00T 2 is

NVIDIA calls GR00T 2 an open “world action model” built on its Cosmos platform. It was trained on robot data and human data, including 100,000 hours of open-source first-person video of people at work. Humanoids Daily describes it as predicting the next few video frames together with the robot’s next actions.

The goal is fine hand skills, such as plugging in cables or aligning gears, learned partly by watching people instead of by costly robot teleoperation.

The reported results

NVIDIA says GR00T 2 completed 40 of 50 evaluated tasks at least once with no task-specific demonstrations. Completing a task once is a lenient bar, so read it as a sign of transfer, not of reliability.

StartupHub reports NVIDIA’s other figures: 71% average success across 50 tasks with one demonstration each, and 95% with five. NVIDIA also says the model ranks in the top three on four benchmarks: RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces. All of these numbers come from NVIDIA.

The training side

Isaac Lab 3.0 lets researchers use AI agents to set up training environments and switch between NVIDIA’s PhysX and Newton physics engines. In one vision-based dexterous-manipulation benchmark, NVIDIA reports up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX. Agents can also call GR00T 2 through an API to run skills in sequence and recover from errors.

NVIDIA names MicroAGI as one user. The company post-trained GR00T 2 with reinforcement learning to swap Ethernet cables with a robotic arm.

What “open” means today

According to Humanoids Daily, NVIDIA’s Jim Fan said the weights for GR00T 2 and a lighter GR00T 2 Turbo will be released in December. The same report says NVIDIA has not yet published the license terms. For now, the model is something to read about, not something to download.

Sources