Nvidia open-sources Flora to make robot video labeling 125x cheaper than GPT-6 Astra
decoder ·
Nvidia has open-sourced Flora, a data engine designed to significantly reduce the cost of labeling robot training videos. The system utilizes three fine-tuned Qwen3.5 models to segment actions, generate descriptions, and evaluate task success, covering all necessary annotation types for robotic learning. Nvidia claims Flora is up to 125 times cheaper than using GPT-6 Astra, estimating costs at 72 cents per hour compared to $88.50. While Flora demonstrates superior performance in detecting action boundaries and identifying mid-episode errors, it trails behind competitors like Gemini 3.8 Flash in pure description accuracy. The tool is already integrated into Nvidia's GR00T model training pipeline, processing vast amounts of video data daily. This release supports Nvidia's broader strategy to dominate the robotics value chain by providing affordable, high-quality training data. By offering an open-source alternative to commercial services, Nvidia aims to keep developers within its hardware ecosystem.
decoder