Kristensen highlights NVIDIA GR00T 2 humanoid test results
NVIDIA GR00T 2 Turbo completed 48 out of 48 popcorn serving trials on a Unitree G1 humanoid, according to results cited by robotics entrepreneur Jens Peder Kristensen in a LinkedIn post. Kristensen, the founder of TinyMobileRobots and Xboarding.ai, described a video showing the smaller model combining whole body control, manipulation and AI reasoning. He also cited a 93% success rate for clearing a table of different objects, attributing the figures to NVIDIA Research.
For onboard operation, Kristensen reported a prediction time of 124 milliseconds for GR00T 2 Turbo running on NVIDIA Jetson Thor inside the G1. That figure describes model prediction, rather than the time needed to complete a physical task. His commentary also distinguished successful research demonstrations from reliable operation over an eight hour factory shift, arguing that real environment assessments should track task success, human interventions and productive operating hours.
The full NVIDIA GR00T 2 model achieved a 95% success rate across 50 manipulation tasks after five demonstrations per task and 30 minutes of fine tuning, according to the NVIDIA results Kristensen cited. He said the model can learn from human videos and demonstrations performed on other robots, transferring skills between different robot configurations. Kristensen identified that transfer capability as a potential way to reduce the time and cost of teaching robots new tasks.
