NVIDIA opens Flora, its GR00T data engine, labelling robot video about 125x cheaper than GPT-6

NVIDIA's GEAR robotics research group released Flora, the data engine it uses to train its GR00T humanoid foundation models, as three open-weight models with a paper and project page announced on Oct 20: Flora-TempSeg (9B) cuts manipulation video into timed actions, Flora-Captioner (27B) writes an instruction for each segment, and Flora-Rewarder (27B) scores progress, success and the reason for a failure.

The group said Flora labels 200,000 hours of video a day on 512 GB300 GPUs and, by its own estimate, costs about $0.72 per hour of source video on eight GB300s at $10 per GPU-hour, against $88.49 for GPT-6 Astra at maximum reasoning effort, roughly 125 times less, while scoring 57.1% versus 53.7% on an internal 50-episode quality test.

Public benchmarks are mixed: Flora leads on WGO-Bench segmentation (F1 50.3 vs 46.0) and on RoboReward error (MAE 40.0 vs 55.4), but trails Astra on gold-window caption accuracy (64.7 vs 77.4). The weights on Hugging Face, built on Qwen3.5, carry NVIDIA's non-commercial licence, so they are for research only and not for production robots or as a sole safety monitor.

MEKAWIRE context: a day earlier NVIDIA unveiled its Cosmos Data Factory at GTC Berlin ([story](/news/nvidia-cosmos-data-factory-gtc-berlin-anybotics-sereact-zalando-ceva-siemens-world-model-tools-20261021)), and Dream Labs, founded by former GR00T researchers, raised a $90M seed on Oct 19; background in our [NVIDIA robotics strategy explainer](/news/nvidia-robotics-strategy-explained-2026-jetson-thor-gr00t-cosmos-free-models-sell-data-center-gpus-20-robot-startups-backed-20261017) and on the [NVIDIA company page](/companies/nvidia).