NVIDIA details GR00T 2 robot demonstrations and results

NVIDIA describes GR00T 2 as an open frontier world action model built on NVIDIA Cosmos[1]. The model predicts future video and robot actions together[3]. NVIDIA says its training used diverse robot and human data, including 100,000 hours of open-source human egocentric data[1]. The results reported for the model span dexterous handling, assembly, and industrial tasks[3]. For teams assessing the research, the range of task examples offers more context than a single headline success rate.

NVIDIA reports that GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations[1]. With one demonstration per task, NVIDIA reports 71% average success across the 50 tasks; with five demonstrations per task, it reports 95%[1]. The research page says the five-demonstration setup uses under three minutes of robot data and a 30-minute fine-tune on one RTX 5090[3]. Other reported examples include a robot preparing banana yogurt and scooping granola, and 85% success seating GB300 memory modules[3].

The examples include both practical manipulation and repeated assembly. NVIDIA reports a 65% success rate for lining a lab funnel with filter paper and wetting it with a wash bottle[3]. It also reports that the YAM arm assembled a toy dump truck end to end in 60% of trials[3]. For three toys, YAM completed at least 97 of 100 back-to-back assemblies for each[3]. NVIDIA says YAM followed what-and-where instructions 92% of the time, compared with 40% using the task instruction alone[3]. These results describe distinct tasks and setups, not one common benchmark.

NVIDIA says GR00T 2 Turbo is compact enough to run on the Unitree G1 humanoid itself[3]. A prediction takes 124 milliseconds on the G1’s Jetson Thor, according to NVIDIA[3]. NVIDIA reports that Turbo’s popcorn-serving task succeeded in all 48 trials through the SONIC whole-body controller, and that it cleared a table of diverse objects in 93% of trials[3]. Separately, NVIDIA says Foxconn is applying GR00T 2 to factory automation for server manufacturing, Siemens is evaluating it across industrial use cases, and Nexuni is testing it for robotic laundry handling[1].