Nvidia opens Flora to cut robot-video labeling costs
Nvidia has open-sourced Flora, a robotics data engine that labels 200,000 hours daily and claims a 125-fold cost advantage over GPT-6 Astra.
What happened
Nvidia has released Flora, an open robotics data engine designed to turn long robot and human-demonstration videos into structured training data. The system combines temporal segmentation, captioning and reward-oriented labeling, allowing researchers to identify actions, sub-tasks and outcomes inside video rather than treating each clip as an undifferentiated sequence.
Nvidia says Flora was used in the data pipeline for its GR00T robotics models and can process about 200,000 hours of data per day on 512 GB300 GPUs. On Nvidia’s internal comparisons, it is competitive with GPT-6 Astra on several annotation tasks while costing roughly $0.72 to label an hour of robot video, versus about $88 through the proprietary alternative.
Why it matters
Robotics development is constrained less by the supply of raw video than by the expense of converting that video into reliable action labels. A cheaper annotation layer could make it practical for universities, smaller robotics companies and open research groups to build specialized datasets instead of depending on a few large labs’ curated collections.
The release also shifts competition toward the data-engineering layer. Flora is not a robot controller, and its outputs still require validation before being used for safety-critical training. Its licensing terms and the quality of labels on unfamiliar embodiments may limit commercial adoption. Even so, publishing the pipeline gives researchers a reproducible starting point for physical-AI data work and could lower one of the least visible costs in robot learning.
Uncle Cat take
Nvidia’s headline number is less important than the 200,000-hour production scale: if outside labs can reproduce even a fraction of it, robotics data stops being a boutique resource.