Nvidia’s Flora Cuts Robot Labeling Costs
Nvidia’s new Flora system labels robot-training videos for far less than GPT-6 Astra, but it doesn’t perform better on every test.
Illustration made on Lumeta with GPT Image 2.5Make one like it
Nvidia has open-sourced Flora, a system for labeling robot training video, and on Oct. 25, 2026 the company said it can do work that might cost about $88.49 per source-video hour with GPT-6 Astra for about $0.72 instead. That matters to robotics teams training on huge video datasets, although Nvidia's price gap is based on estimates rather than production billing, and researchers can try the released models now on Hugging Face.
What changed
Flora is designed to produce three kinds of labels from robot video: timestamped action descriptions, step-by-step progress ratings and a final success judgment for the whole task. Nvidia says the system is already used in training its GR00T models, processing up to 200,000 hours of video per day across 512 GB300 GPUs. Under the hood, Flora uses three fine-tuned Qwen3.5 models, including a 9B segmenter and two 27B models for captioning and reward scoring.
Illustration made on Lumeta with GPT Image 2.5Make one like it
What the price claim means
In Nvidia's 50-video test, Flora scored 57.1 percent on annotation quality versus 53.7 percent for GPT-6 Astra, while Nvidia estimated the cost at $0.72 per video hour for Flora and $88.49 for Astra. The paper says those Flora costs are projections built from an assumed GPU rental price of $10 per GPU hour, not measured deployment bills. Nvidia also reports Claude Opus 5 at $41.70 per video hour and Gemini 3.8 Flash at $8.48 in the same comparison.
Where Flora seems strongest, and where it does not
Nvidia's results suggest Flora is especially good at finding action boundaries, spotting failed steps inside an episode and giving partial credit when a robot makes some progress without fully finishing a task. On RoboRewardBench, Nvidia says Flora's error rate is 27.8 percent lower than Astra's, and Flora-Rewarder catches 53.3 percent of failed steps while Astra catches none in that test. But Astra still leads on some measures, including overall WGO-Bench score, robot-video results on Nvidia's DenseCaptionBench, rubric-based scoring and trajectory ranking. That leaves Flora looking less like a total replacement for frontier multimodal models and more like a specialized open option for labs that need cheap, high-volume robotics annotation at scale.
