NVIDIA releases GR00T 2, an open model for robot hand tasks
With one demonstration for each task, GR00T 2 gets 71% average success on 50 tasks, and 95% with five.
At the Conference on Robot Learning in Austin, Texas, NVIDIA released GR00T 2, an open world action model. This type of model shows how its moves change the scene. It is on NVIDIA Cosmos and it is for robot tasks with cables and gears. NVIDIA tells of 71% average success on 50 tasks with one demonstration for each task.
How it works
GR00T 2 uses robot data and human data, with 100,000 hours of human first-person video. It moves skills between different robot bodies. AI agents can use its software connection to do skills in sequence and to correct errors.
The numbers
- It did 40 of 50 tasks one time or more with no task demonstration.
- Average success is 71% with one demonstration and 95% with five.
- Isaac Lab 3.0 with Newton has a training speed of a maximum of 2.5 times that of Isaac Lab 2.3.
What is not known
These results are from NVIDIA. The source gives no tests by other persons. Nexuni, Foxconn and Siemens test the model, but the source gives no results from them.
Insight: Open robot tools can pull makers into one simulator stack
GR00T 2, Isaac Lab and Newton are open, and Skild AI made its STEEL solver on Newton. A study of 6 November found that learning curve checks found one silent simulator fault in ten tests at maximum. This reads as a toll gate: free parts that pull robot makers into one simulator stack, where the simulator has a small number of checks.
Sources
Posted
