NVIDIA GR00T 2 transfers human task skills to humanoid robots
NVIDIA has introduced GR00T 2, a robot foundation model that the company says can transfer human demonstrations to humanoids and other robot bodies, according to Interesting Engineering. A faster variant, GR00T 2 Turbo, runs on a Unitree G1 humanoid and makes predictions in 124 milliseconds using onboard computing, the outlet reports. The model jointly predicts future video and robot movements, linking how a physical interaction is expected to unfold with the actions needed to carry it out.
Across 50 tasks that had not been demonstrated on the target robot, NVIDIA reports an average success rate of 40 percent without training specific to those tasks. Five demonstrations per task raised the reported average to 95 percent after a 30 minute fine tuning session on a single NVIDIA RTX 5090 graphics card. Those demonstrations amounted to less than three minutes of robot data per task. The distinction is important for interpreting the result: the 95 percent figure describes performance after adaptation, while the 40 percent result measures transfer without that additional task training. The evaluation covers different robot bodies, rather than being presented as a humanoid only benchmark.
NVIDIA calls its transfer approach X-Adapt. According to Interesting Engineering, GR00T 2 learns from 402,730 hours of human video alongside handheld gripper and robot teleoperation data, and is designed to support parallel grippers, dexterous five finger hands and humanoids. In one experiment, a human demonstrated lining a laboratory funnel with filter paper and wetting it with a wash bottle, without teleoperating the target robot. NVIDIA reports that a robot with five finger hands then completed the task in 65 percent of trials. That experiment tests a particularly useful capability for humanoid manipulation: learning a task demonstrated by a person rather than requiring the same task to be demonstrated through the robot itself.
Sources: Interesting Engineering, breakingthenews.net
