NVIDIA’s GR00T 2 robot-control model averaged 40% success across 50 tasks before seeing demonstrations on the target robot, then 95% after five demonstrations per task. The result offers a striking example of robots learning from limited task-specific examples, with important qualifications about how the tests were chosen.
NVIDIA Watch analysis
What happened
A report published by R&D World on 11 November describes NVIDIA’s GR00T 2 experiments. The model uses video, language instructions and information about the robot’s configuration to generate predicted video alongside movements for the robot to execute. NVIDIA says the model completed 40 of 50 tasks at least once before task-specific demonstrations; after five demonstrations per task, it succeeded in every trial on 37 tasks.
The headline result needs its footnote: NVIDIA selected tasks on which GR00T 2 could reach at least 80% success after five demonstrations. The paper says that makes the high post-demonstration result “high partly by construction”. The article also reports separate chemistry-lab tests with Dexter, a two-armed robot fitted with Sharpa hands. With human demonstrations and data from other robots, Dexter completed a filter-paper setup in 13 of 20 trials, but managed only 30% on dispensing liquid and 10% on slide dabbing. Adding 10 demonstrations using the target robot raised those results to 95%, 70% and 50%, respectively.
Why it matters
The useful advance is not a robot passing a single party trick. It is evidence that combining human video and experience from other robots can help a system adapt to tasks on a specific robot, with relatively few additional demonstrations. That could make teaching robots new procedures less dependent on collecting large amounts of task-specific robot data.
But the results are NVIDIA’s evaluations, not an independent comparison across every robot or task. Some tests remain far from reliable: even after additional training, Dexter succeeded in only half of the slide-dabbing trials. And task selection matters. A headline percentage can look wonderfully tidy while the experimental setup does some of the tidying.
Our read
This is a substantive robotics result, and the strongest detail is the combination of broad prior training with a small number of demonstrations for a particular task. The chemistry-lab results make the promise more concrete, and the uneven scores make the limitations just as concrete. The next question is whether the approach transfers to new tasks and settings without carefully selected demonstrations.
What to watch
- Whether independent teams reproduce the results on different robots and tasks.
- How performance changes when tasks are not selected for strong post-demonstration results.
- Whether fewer task-specific demonstrations translate into practical time or cost savings for robot operators.
Discussion spark: For teaching robots new jobs, would you prioritise strong results on a selected set of tasks, or reliable performance across a broader and less curated test?
Sources and evidence
- Source update (11 November 2026, 21:48 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.