The News, Explained

NVIDIA introduced GR00T 2 at the Conference on Robot Learning, or CoRL, on November 9, 2026. The robot foundation model learns manipulation skills such as grasping objects and inserting components. One headline result is 95% average success after five additional demonstrations per task. Understanding that number requires looking at the robot, the tasks and the practice behind it. Source Source

GR00T 2 is a world-action model: it predicts future video and robot actions together. It takes recent video, an instruction and the robot state, generates a future scene and a chunk of actions, executes the actions and repeats. That structure learns how a scene changes as the robot acts. Pretraining uses human activity video and manipulation data from multiple robots. Source

The adaptation experiment covers 50 tasks on a bimanual robot with two Sharpa hands. The tasks were absent from the target robot’s own training data, but appeared in human video or other robots’ data. A new task here means that the target robot had not demonstrated that task. The model still brings pretraining and experience from other bodies to the evaluation. Source

Without additional task demonstrations, the model completed 40 of the 50 tasks at least once. Average success across trials in that condition was 40%. The first measure describes the breadth of tasks with a successful attempt; the second describes how often attempts succeed. Forty tasks and a 40% success rate have different denominators. Source Source

Across the same 50 tasks, the researchers report 71% average success after fine-tuning on one demonstration per task, rising to 95% with five. Fine-tuning adapts an already trained model using additional data. Five demonstrations supplied under three minutes of robot data per task. Each adaptation used a 30-minute LoRA fine-tune on a single RTX 5090. LoRA trains a small set of additional model parameters. Those 30 minutes describe adaptation, rather than the preceding full pretraining process. Source

The 95% figure is an NVIDIA research result for the specified robot and 50 tasks. With five demonstrations, 37 tasks succeeded in every evaluation trial. Reading task-level results alongside the average helps distinguish the range of skills from consistency across attempts. NVIDIA also describes laundry-handling tests, server-manufacturing applications and industrial evaluations, without assigning that same success rate to every deployment. Source Source

Availability has a separate timeline. The NVIDIA research page marks the code as available in December. Reading the published research and demonstrations today and reproducing the result with that code on another robot are separate stages. The announcement covers model research, bounded evaluations and industrial tests or applications. Source Source

OYOPICK’s Take

Improving a target robot’s performance with a few additional demonstrations is encouraging because it could reduce the effort needed to teach new work. Transferring experience from people and other robots to a new body points toward robots helping with a wider range of activities. We would judge that potential through the required demonstrations, the target robot’s training history and repeated attempts, alongside the average. The distinction between 40 successful tasks and 40% success shows why those details matter. Source

If the December code release follows, we hope more researchers can examine these conditions on other tasks and in other environments. Comparing the conditions behind failures and the process of recovering from them, alongside successful demonstrations, would make that work more useful. Whether the reported results translate into reliability in a deployment remains a question for those follow-up evaluations. Source