Nvidia Unveils GR00T 2, Teaching Robots Tasks from Human Videos
Nvidia has released GR00T 2, a new AI model for robots. It is designed to teach robots tasks without elaborate remote control, by transferring human hand movements from videos onto robotic arms. Reported success rates vary widely depending on the task.
Nvidia has unveiled GR00T 2, a new AI model for robots. The goal is to teach robots new tasks without humans having to laboriously control them through teleoperation. Instead, the model relies on video footage in which people perform the respective activities.
The method works by extracting human hand movements from video recordings and transferring them to robotic arms. The robot thus learns to imitate a movement it has only previously seen on video. This eliminates much of the effort previously required to teach robots step by step by hand.
According to the available information, the foundation consists of around 400,000 hours of video material. The scale of this data basis underscores how strongly development relies on large volumes of observational data to expand the capabilities of machines.
How reliably the system works apparently depends heavily on the specific task. The reported success rates range from 10 to 80 percent. This wide spread shows that GR00T 2 performs considerably better on some activities than on others. A statement applying equally to all applications cannot be derived from this.
The approach aims to expand the range of applications for robots without requiring every new capability to be programmed or remotely controlled with high manual effort. Whether the method proves successful beyond the tasks shown is not clear from the available information.
Source: the-decoder.com
19
