Nvidia unveils GR00T 2 robot AI model and 100,000-hour dataset
The world action model is available now with data and tools; weights and per-robot adaptation methods arrive in December.
Nvidia unveiled GR00T 2, a world action model (WAM) that predicts robot actions together with changes in the surrounding environment, according to digitaltoday.co.kr. A 100,000-hour human demonstration dataset and related tools for the model are available now.
Model weights and per-robot training and adaptation methods are scheduled to be released in December. GR00T 2 can be applied to different robot forms, including robot arms, five-fingered robot hands and humanoids.
In Nvidia's evaluation, GR00T 2 completed 40 of 50 tasks at least once without any demonstration on the target robot, with an average success rate of 40%. Adding one demonstration per task raised the success rate to 71%, and five demonstrations raised it to 95%.
Lightweight model for humanoids
GR00T 2 Turbo, a lightweight model for humanoids, runs on Jetson Thor. Foxconn is applying GR00T 2 to automate its server manufacturing factories.
Siemens is evaluating GR00T 2 for industrial use, and Taiwan robot company Nexuni is testing the model for laundry handling tasks.
Quick answers
When will the GR00T 2 model weights be released?
Model weights and per-robot training and adaptation methods are scheduled to be released in December. The 100,000-hour human demonstration dataset and related tools are available now.
What robot forms can GR00T 2 be applied to?
Nvidia says the model can be applied to robot arms, five-fingered robot hands and humanoids.
What success rate did GR00T 2 show in Nvidia's evaluation?
It completed 40 of 50 tasks at least once without demonstration on the target robot, at an average success rate of 40%. One demonstration per task raised this to 71% and five demonstrations to 95%.