Back to news
launchNVIDIAAlibaba2026-10-26

NVIDIA open-sources Flora robot video labeling engine at $0.72 per hour

NVIDIA's GEAR lab open-sourced Flora on Oct 26, a data engine for labeling robot video used in training. The system is built on three models fine-tuned from Alibaba's Qwen3.5, with a 9B model segmenting action clips, a 27B model generating action instructions and a 27B reward model scoring them, covering action boundary detection, step-by-step execution scoring and task success judgment in one pipeline, together with an in-house action segmentation model and the Flora-Rewarder. NVIDIA estimates labeling at about $0.72 per hour against roughly $88.50 per hour with GPT-6 Astra, a gap of 122 to 125 times, a figure derived from modeling GPU rentals at $10 per hour per GPU rather than measured in production. In a 50-video test Flora scored 57.1 percent overall against 53.7 percent for GPT-6 Astra, leading on temporal segmentation and human-recorded footage but trailing Gemini 3.8 Flash on pure description accuracy and ranking second on robot video description, with Astra also serving as evaluator in the same test. Weights are on Hugging Face, and internally NVIDIA labels about 200,000 hours of robot video per day with Flora, which already serves as the data engine for training GR00T on 512 GB300 GPUs.

NVIDIAFlora机器人标注具身智能Qwen3.5开源