Nvidia open-sources Flora for cheaper robot video labeling

Nvidia has open-sourced Flora, a data engine designed to lower the cost of labeling robot video. The company states that using Flora to label one hour of video costs 72 cents. This figure is significantly lower than the roughly $88.50 cost for GPT-6 Astra and the $42 cost for Claude Opus 5. Nvidia calculated these figures based on an assumed GPU rental price of $10 per hour.

Flora consists of three fine-tuned Qwen3.5 models. A 9B model splits videos into action segments, while a 27B model describes actions with context. A component called Flora-Rewarder rates episode success on a scale of 1 to 5 and classifies half-second steps as normal, suboptimal, or failure. The system is currently running during the training of Nvidia's GR00T models, annotating up to 200,000 hours of video per day across 512 GB300 GPUs.

On the public WGO-Bench, Flora detects action boundaries best but trails Astra in overall score. It also falls behind Gemini 3.8 Flash in description accuracy. On Nvidia's DenseCaptionBench, Flora leads all metrics for human recordings but ranks second behind Astra for robot videos. On RoboRewardBench, Flora's error rate is 27.8 percent lower than Astra's. Flora-Rewarder catches 53.3 percent of failed steps, a rate where no competitor model exceeds 11.7 percent.

The Flora models are available on Hugging Face.