Transferring human skills to robots with GR00T 2

Trained on human video and robotics data, GR00T 2 is tested by NVIDIA on skill transfer between machines and adaptation from a few demonstrations.

Across fifty tasks absent from the target robot’s demonstrations, GR00T 2 achieves a 40% success rate without task-specific adaptation. Five demonstrations per task raise that average to 95%, following approximately thirty minutes of fine-tuning on an RTX 5090. These internal NVIDIA results illustrate the model’s approach: reuse skills observed in humans or other machines, then adapt their execution to the robot performing the task.

The scope of the experiment matters as much as the score. The technical report specifies that the fifty tasks were selected from those the model learns with five demonstrations, achieving at least 80% success. The high final score therefore partly reflects that selection. Before adaptation, forty of the fifty tasks are completed at least once: that count does not correspond to an 80% average success rate across trials. Nor does the 95% figure describe a model capable of performing any task after three minutes of observation: the demonstrations and fine-tuning apply separately to each task.

Presented at the CoRL conference, the project belongs to the World Action Model family, predicting actions and the scene’s visual evolution together. Code is scheduled for release in December 2026. The research page does not yet specify licensing terms for the model, which NVIDIA describes as open.

The system uses recent footage from three cameras, an instruction, and the robot’s state. A diffusion transformer jointly produces future video and a chunk of action commands. Components specific to each machine translate the outputs into actions suited to its body. The robot executes a sequence, collects new observations, and repeats. The future videos shown on the project page are model predictions, distinct from footage actually recorded during execution.

This rapid adaptation follows substantial pretraining. The described mixture includes 402,730 hours of first-person human video, 10,184 hours of demonstrations using a handheld gripper, and 23,500 hours of robot teleoperation. The data covers gripper-equipped arms, dexterous hands, and a humanoid. A subset containing 100,000 hours of human video is available on Hugging Face. The few demonstrations added for a new task build on that prior training.

Transfer from humans is also tested separately. On Dexter, a robot equipped with two Sharpa hands, placing filter paper in a funnel achieves 65% success over twenty trials, although the task was demonstrated only by people. The robot nevertheless has general data teaching it how its own body looks and moves. Results vary substantially by action: another laboratory manipulation task, dabbing a microscope slide, succeeds in only 10% of trials.

The distinction between decision-making and execution appears in experiments involving a separate reasoning agent. That agent breaks down the task, monitors progress, and calls GR00T 2 through an API for manipulation. It can also use other control tools to retrieve a dropped part or recover from a stall. In this configuration, the YAM arm assembles three toys in 98, 97, and 97 out of one hundred trials, respectively. Those scores apply to the combined agent, model, and recovery tools; the adapted policy without that agent achieves between 71 and 84 successful assemblies, depending on the toy.

A second control method adds marks to the camera image indicating where to grasp and release an object. On YAM, detailed instructions combined with these markers raise success on short placement instructions to 92%, compared with 40% using only the overall task instruction. For a longer waste-sorting task, the same method achieves seventeen successful runs out of twenty. Results on Dexter remain more limited: the markers help move more objects but rarely lead to completion of the entire sorting task.

The Turbo variant targets onboard execution on the Unitree G1 humanoid. Using a more compact video representation, it produces a prediction in 124 milliseconds on Jetson Thor, according to the published measurements. That time covers computation of a motion sequence. The SONIC controller then translates whole-body motion commands into joint targets at 50 Hz. The high-level reasoning model remains outside this onboard execution stack.

The G1 trials also show that success still depends on the task. The robot completes all forty-eight popcorn-serving trials and clears a table in approximately 93% of the thirty trials devoted to that task. Cleaning up a liquid spill remains at approximately 47% across thirty trials.