CONFIRMED: NVIDIA has opened Flora, the data engine from its GEAR robotics lab used to train its GR00T humanoid models. The suite turns robot and human demonstration video into training labels: where each action starts and ends, what was done, and whether the execution was good. The models are published on Hugging Face under a non-commercial research licence.
There are three models, each with a distinct job. Flora-TempSeg, with 9 billion parameters, splits video into action intervals and marks the start and end of each step. Flora-Captioner, with 27 billion parameters and post-trained from Qwen3.5-27B, writes one short instruction per interval, taking into account what happened before and after. Flora-Rewarder, also with 27 billion parameters, gives an episode success score from 1 to 5, classifies short steps as normal, suboptimal or failed, and assigns scores against user-defined rubrics.
The scale claimed by NVIDIA is the most striking part of the announcement. According to the company, Flora already labels up to 200,000 hours of video per day on 512 GB300 GPUs and was the data engine used to train GR00T-2. That figure is the manufacturer's own claim, not an independent measurement, and it describes labelling throughput — not newly collected hours of robot experience.
Cost is the second central promise, and the one that most needs careful reading. By NVIDIA's estimate, labelling one hour of video costs about $0.72 with Flora, versus about $88.49 with GPT-6 Astra at maximum effort — a difference of roughly 125 times. The calculation assumes GPU rental at $10 per hour and covers the captioning step, not the reward model, and it has not yet been verified in production by third parties.
On quality, the picture is mixed, according to NVIDIA's own internal evaluation on 50 episodes. Flora scored 57.1% against Astra's 53.7% on the overall annotation metric, with an advantage on human videos, action-boundary detection and partial-success cases. Astra led on robot videos, description accuracy and rubric scores. On a public action-boundary test, Flora also led on temporal localisation, but trailed Astra and Gemini 3.8 Flash on overall description accuracy. The material itself acknowledges limits: Flora sometimes splits long actions into too many steps and can describe an action incorrectly.
For researchers, the practical limit is the licence. The models are governed by NVIDIA's OneWay non-commercial licence, intended for research and development, with no support for commercial or production use. The Flora-Rewarder card also makes clear that its outputs are annotations: the model should not directly control a robot or be the sole safety monitor of a physical system.
Flora's value, if the numbers hold outside NVIDIA's lab, lies in unlocking robotics' least visible bottleneck: filming demonstrations is not enough; hours of video must be turned into usable training data. Smaller labs could build larger datasets without paying a frontier model's price per hour of video. Until then, the test that matters is the one NVIDIA has not yet shown independently: running Flora on your own videos, measuring segmentation and description errors, and comparing the real cost per hour.
Sources: Flora-Rewarder model card on Hugging Face (https://huggingface.co/nvidia/Flora-Rewarder); The Decoder, “Nvidia open-sources Flora to make robot video labeling 125x cheaper than GPT-6 Astra” (https://the-decoder.com/nvidia-open-sources-flora-to-make-robot-video-labeling-125x-cheaper-than-gpt-6-astra/); Tesla News Wire, “Nvidia Flora: open-source robot video labeling for GR00T” (https://teslanewswire.com/articles/nvidia-flora-open-source-robot-video-labeling-models-gr00t-200000-hours-oct-2026).
Image credit: Flora capabilities graphic, an NVIDIA image reproduced by The Decoder.
Transparency: this article was prepared with AI assistance, from cross-checking public sources on the internet and public posts on X. NVIDIA's cost and performance claims still await independent replication. Consult the original sources to confirm details before any technical decision.
Muse
