Nvidia open-sources Flora, a three-model video labeling engine for robot training data, with self-reported cost of $0.72 per video hour

Nvidia's GEAR lab released Flora, an open set of three fine-tuned Qwen3.5 models (a 9B action segmenter, a 27B captioner and a 27B reward model called Flora-Rewarder). Together they turn robot video into action descriptions, step-by-step progress ratings and task success judgements. The weights are on Hugging Face and a paper is published. Nvidia says Flora already runs in GR00T training, labelling up to 200,000 hours of video a day on 512 GB300 GPUs.

Robotics teams short of labelled video now have an open, inspectable labelling pipeline they can test and adapt instead of paying a commercial service or a frontier model API. That mainly helps researchers and developers with their own GPU capacity, and it puts some price pressure on paid data services such as Scale AI's. The cost advantage is unproven for anyone outside Nvidia, and it depends on having serious compute, so the practical benefit for small teams is uncertain.

The "125x cheaper" headline rests on a projection that assumes $10 per GPU-hour, not on measured spend, and the quality gap (57.1% vs 53.7%) is small and comes from Nvidia's own evaluation. GPT-6 Astra also acted as a judge in that evaluation, and Astra still wins on several benchmarks. The comparison also sets a purpose-built labeller against a general model, which flatters Flora. The story does not say how well the labels improve the robots trained on them.

1. 01Independent replications of the $0.72 per hour cost and the 57.1% vs 53.7% quality result, ideally with a different judge model and real (not assumed) GPU pricing
2. 02Evidence that GR00T or other robot policies trained on Flora labels perform better on downstream tasks, not just on annotation benchmarks
3. 03Adoption signals over the next one to three months: Hugging Face downloads, third-party fine-tunes, and any response or price change from Scale AI or other labelling vendors