Nvidia projects open Flora cuts robot video labeling 125x
Nvidia claims robot video labeling just got 125x cheaper.
Nvidia says its open data engine Flora can label an hour of robot video for about 72 cents, against roughly $88.50 for GPT-6 Astra.
That is the 125x gap in the headline. It is a projection, not a production measurement. Nvidia calculated it from an assumed rental price of $10 per GPU per hour.
Robot models learn from video that needs three kinds of labels: what action happens when, how well a robot is progressing step by step, and whether a task succeeded. Most existing tools provide only one. Flora covers all three. Nvidia says it is already running during training of its GR00T models, annotating up to 200,000 hours of video per day.
In a 50-video test, Flora scored 57.1 percent annotation quality versus 53.7 percent for GPT-6 Astra. Its Flora-Rewarder model catches 53.3 percent of failed steps, but it falsely flags about 11 percent of normal steps.
The other benchmarks are split. Astra still leads on rubric scoring and trajectory ranking, and Astra also judges some of the tests it is part of, which the researchers acknowledge. The models are on Hugging Face.
Does a 72-cent estimate change how you see robot training?
Archive video: Nux, CC BY 4.0 (Wikimedia Commons); NASA
Video: research.nvidia.com/labs