Nvidia introduced GR00T 2 at the Conference on Robot Learning in Austin on November 9, presenting a robot model intended to adapt to unfamiliar manipulation tasks with limited demonstrations. The company’s research page says its code will become available in December.
The announcement puts the emphasis on handling objects rather than simply moving through a room. Nvidia’s accompanying blog describes applications under investigation that range from laundry handling to server manufacturing, alongside new simulation and evaluation tools.
Learning actions and their consequences
Nvidia Research describes GR00T 2 as a world action model, a system that predicts future images together with the robot movements that would produce them. Its training combines human video, handheld-gripper demonstrations and robot operation data.
In company tests covering 50 tasks not previously demonstrated on the target robot, Nvidia reports a 40 percent success rate without task-specific adaptation. Five demonstrations per task increased that figure to 95 percent after a 30-minute adjustment on one RTX 5090 graphics card.
The distinction between tasks attempted and trials completed matters. Nvidia’s blog separately says the model finished 40 of the 50 tasks at least once without task-specific demonstrations. That measures the breadth of tasks with some success, rather than an 80 percent success rate across every attempt.
From isolated skills to longer jobs
The research page also describes combining GR00T 2 with a reasoning model that chooses steps and checks progress. In three toy-assembly tests, the combined system completed at least 97 of 100 trials for each toy, according to Nvidia.
That division of work illustrates an important engineering question: can a robot identify a failed step and recover without restarting the entire job? For an employer considering automation, recovery time and the frequency of human intervention can matter as much as the best successful demonstration.
Nvidia says Nexuni is testing laundry handling, Foxconn is applying the model to server-manufacturing automation and Siemens is evaluating industrial uses. These descriptions identify different stages of work. The announcement does not provide evidence of a common production rollout or independently measured factory savings.
What developers still need to establish
The company is also showcasing Isaac Lab 3.0 and its integration with the Newton physics engine. Nvidia says these tools support training with objects such as cloth and cables before moving a learned skill onto physical equipment.
Simulation can help organize development, but buyers should still ask for tests on their actual objects, lighting, equipment and safety requirements. A result obtained with one robot hand or a carefully arranged task should not be assumed to transfer unchanged to a different workplace.
The immediate milestone is the planned December code release. Independent replication, the accompanying license and the requirements for adapting the model to specific hardware will be important checks before the research results can be treated as evidence of dependable commercial performance.
