Nvidia open-sources Flora to make robot video labeling 125x cheaper than GPT-6 Astra

Why it matters

Nvidia has released Flora, an open-source data engine that labels robot training videos at a fraction of the cost of commercial alternatives like GPT-6 Astra. The system, built from three fine-tuned Qwen3.5 models including Flora-Rewarder, handles all three annotation types robotics training needs: action descriptions, step-by-step progress ratings, and task success judgments. Most existing tools cover only one of these, while general-purpose language models are too expensive for the huge volumes of video involved. Flora already annotates video daily during training of Nvidia's GR00T robot models, and the models are available on Hugging Face. The benchmarks are mixed: Flora excels at detecting action boundaries and catching failed steps mid-episode, where competitors barely register, but trails Astra on description accuracy and rubric-based scoring. Nvidia also acknowledges a methodological wrinkle, since Astra both evaluates and competes in some tests. The release fits Nvidia's broader strategy of open-sourcing AI tools to keep developers in its hardware ecosystem and offset the shortage of robotics training data, with Google's Gemini 4 Argon targeting similar territory.