Nvidia releases GR00T 2 world-action model, robots can learn skills from human videos
Nvidia released GR00T 2 at the Conference on Robot Learning held in Austin on November 9, describing it as a new generation of foundational world-action models. The model predicts future video and robot actions at the same time, and can run on parallel grippers, dexterous hands, and humanoid robots.
The company said GR00T 2 can learn skills from human videos, handheld gripper demonstrations, and other robot data, and transfer those skills to target robots that have never demonstrated the task. On 50 tasks unseen by the target robots, the zero-shot average success rate was about 40%; after providing 5 demonstrations for each task (less than 3 minutes of robot data) and fine-tuning on a single RTX 5090 for 30 minutes, the average success rate rose to 95%.
The model code is expected to be released in December. The pretraining data include about 402,700 hours of first-person human video, 10,200 hours of handheld gripper demonstrations, and 23,500 hours of robot teleoperation data.
Original source:
Title: GR00T 2: Scaling Robot Foundation Models for Generalization, Dexterity, and Adaptation
Publisher: NVIDIA GEAR Research Team
Date of publication: November 2026
URL: research.nvidia.com/labs/gear/gr00t-2
Source type: company research page
Verification status: the full page was actually read for this report
Supporting content: GR00T 2 is a world-action model that can learn from human videos and other robot data and transfer skills; zero-shot success rate on 50 tasks is about 40%, rising to 95% after 5 demonstrations
ABAB AI Insight
Nvidia released GR00T 2 at the Conference on Robot Learning on November 9, 2026, positioning it as a world-action model rather than the earlier vision-language-action model. The model uses a diffusion transformer to predict future video frames and action chunks at the same time, thereby learning from the visual results of physical interaction.
The company disclosed that the pretraining data are dominated by first-person human video (about 402,700 hours), supplemented by handheld gripper demonstrations and robot teleoperation data. The official page states that skills can be transferred from videos consisting only of human demonstrations to the target robot, without prior demonstrations of the task by that robot.
On 50 tasks unseen by the target robots, Nvidia reported a zero-shot average success rate of about 40%, rising to 95% after 5 demonstrations per task. These figures come from the company’s self-reported tests; the page does not list all task names, trial counts, or independent replication results. The model weights and code are expected to be released in December.
GR00T 2 also offers a Turbo variant that can run on the onboard computing of the Unitree G1 humanoid robot; it also supports cooperation with reasoning models to handle multi-step tasks and error recovery through an API.
ABAB News · Cognitive Law
Nvidia released the GR00T 2 world-action model at CoRL on November 9, predicting future video and robot actions at the same time.
The company said the model can learn skills from human videos and other robot data, and transfer them to target robots that have not seen the task before.
On 50 tasks, the zero-shot average success rate was about 40%, rising to 95% after 5 demonstrations per task; the code is expected to be released in December.