AI News

NVIDIA Releases World Action Model GR00T 2 and Launches Isaac Lab 3.0

English text is being prepared.Updated Nov 10 12:32

AI summary

NVIDIA has released its new-generation world action model GR00T 2, which can simultaneously predict future video frames and robot actions. Its pretraining used over 400,000 hours of first-person human video, about 10,000 hours of handheld gripper demonstrations, and 23,500 hours of robot teleoperation data. Across 50 tasks that the target robot had not been trained on before, GR00T 2 achieved an average zero-shot success rate of 40%; after providing one demonstration per task, the average success rate rose to 71%, and after five demonstrations plus fine-tuning, it reached 95%. NVIDIA also introduced the robot learning framework Isaac Lab 3.0, which integrates the Newton physics engine and achieves up to 2.5 times higher training throughput than previous solutions in vision-driven dexterous manipulation benchmarks.

Why read this

AI score 70

The original text provides zero-shot success rates, the number of demonstrations, and the improvement curve after fine-tuning, which can be used to assess how usable this world action model is on real tasks.

Topics

Tags

Body · AI translation

English text is being prepared.

On November 9, NVIDIA released its next-generation world action model GR00T 2, and introduced the robot learning framework Isaac Lab 3.0 along with multiple robot training and evaluation technologies.

It is reported that GR00T 2 adopts a world action model architecture, capable of simultaneously predicting future video frames and robot actions, enabling robots to execute operations according to instructions, and also to learn the impact of actions on the physical environment.

According to NVIDIA, GR00T 2 was pretrained using over 400,000 hours of human first-person videos, approximately 10,000 hours of handheld gripper demonstrations, and 23,500 hours of robot teleoperation data, and supports transferring skills learned from human demonstrations to different robot embodiments.

In 50 tasks that the target robots had not been previously trained on, GR00T 2 achieved an average zero-shot success rate of 40%, with at least one successful completion in 40 of the tasks. With only one demonstration provided per task, the average success rate rose to 71%; increasing to five demonstrations and after fine-tuning, the success rate further reached 95%.

In terms of fine manipulation, a robot equipped with dual 22-degree-of-freedom dexterous hands has been able to complete tasks such as installing GB300 memory modules and assembling toy trucks, with corresponding test success rates reaching 85% and 65%, respectively. NVIDIA also demonstrated the robot's ability to learn laboratory operations through human demonstrations and to improve performance on insertion tasks using reinforcement learning.

Meanwhile, Isaac Lab 3.0 further integrates the Newton physics engine, supporting simulation of deformable objects such as cables and cloth, as well as complex mechanical joints. In vision-driven dexterous manipulation benchmarks, its training throughput is up to 2.5 times higher than previous solutions.

In terms of industrial applications, Foxconn is applying GR00T 2 to server manufacturing automation, Siemens is evaluating its industrial application potential, and Nexuni is conducting robot clothing handling tests.

Submissions/tips: tougao@youxituoluo.com

Article/business cooperation: Lin Nan (WeChat 19250561593)

Join the industry exchange group: Lin Nan (WeChat 19250561593)

Digital technology industry information and service platform

English text is being prepared.

Source

English text is being prepared.