Robot learning runs on labeled video, and the labeling has been the expensive part. Nvidia’s GEAR lab says it needs three kinds of annotation to train and monitor a machine: when an action happens, how well the attempt went, and whether the task finished at all.
Most tools supply one. Large frontier models can produce all three but at a price that makes bulk footage impractical. Flora, an open data engine Nvidia published this week, runs the whole job on three fine-tuned Qwen3.5 models.
A 9B model slices footage into action segments. A 27B model writes descriptions with the surrounding context in view. A second 27B model, Flora-Rewarder, does three jobs from one prompt: scoring an episode from one to five, tagging each half-second step as normal, suboptimal or failed, and grading against rubrics it writes itself.
The cost gap is the headline. Nvidia estimates about 72 cents to label an hour of video, against roughly $88.50 for GPT-6 Astra and about $42 for Claude Opus 5. Those figures are projections based on an assumed rental rate of $10 per GPU hour, not metered production bills.
On 50 test videos Flora edged out Astra at 57.1 percent annotation quality versus 53.7 percent. It trails on the public WGO-Bench, where it finds action boundaries best but loses on overall score and on pure description accuracy to Gemini 3.8 Flash.
Nvidia says Flora already annotates up to 200,000 hours a day across 512 GB300 GPUs while training its GR00T models. The weights are on Hugging Face.
