Nvidia has released Flora, an open-source data engine designed to annotate robotics training videos. The system utilizes three fine-tuned Qwen3.5 models to provide action descriptions, step-by-step progress ratings, and task completion assessments.
According to Nvidia, Flora can label one hour of video for approximately $0.72, compared to $88.50 for GPT-6 Astra. These cost projections are based on an assumed GPU rental price of $10 per hour.
The system is currently used to train Nvidia's GR00T models, processing up to 200 000 hours of video daily across 512 GB300 GPUs. The models are now available on the Hugging Face platform.
In internal tests, Flora demonstrated a 27.8 percent lower error rate than Astra on the RoboRewardBench. While Flora excels at temporal localization and identifying action boundaries, it occasionally misidentifies specific actions.
Nvidia positions Flora as an open-source alternative to commercial data labeling services. The release is part of the company's broader strategy to support its hardware ecosystem and expand its robotics value chain.
