By Hamza Faisal
Nvidia has open-sourced Flora, the data engine its GEAR robotics lab uses to turn raw robot and human manipulation video into the labels that train its GR00T humanoid models. The headline claim is striking: labelling one hour of video costs about $0.72 with Flora, compared with roughly $88.49 using OpenAI's GPT-6 Astra at maximum effort — a gap of about 125 times.
The system comes from a research paper dated October 20, and its models are available on Hugging Face for non-commercial research. Nvidia says Flora is already battle-tested at scale, labelling 200,000 hours of video a day across 512 of its GB300 chips, and that it was the data engine used to train GR00T-2.
Three models, three kinds of labels
Robot training video needs three kinds of annotation, and most existing tools provide only one. Flora provides all three. Flora-TempSeg, a 9-billion-parameter model, splits a video into action intervals, marking where each step begins and ends. Flora-Captioner, at 27 billion parameters, writes one short instruction per interval — for example, "Place the blue mug on the tray with your right hand" — taking the surrounding context into account. Flora-Rewarder, also 27 billion parameters, scores each episode: an overall success rating from 1 to 5, a classification of every half-second step as normal, suboptimal or failed, and scores against rubrics the user writes.
Those reward signals are what let a robot learn by trial and error. Flora-Rewarder catches 53.3 percent of failed steps inside an episode while no competing model gets above 11.7 percent — and Astra catches none. In one example, Flora identified 34 of 37 individual actions in a video full of short hand movements, while Astra caught only four.
The 125-times-cheaper claim — and its caveats
The cost comparison deserves a careful read. In a 50-video test, Flora scored 57.1 percent on annotation quality against Astra's 53.7 percent, with Astra allowed its full reasoning capacity. But the $0.72 figure is a projection, not a production measurement: Nvidia calculated it from an assumed rental price of $10 per GPU-hour. Real-world costs will vary with hardware, utilisation and setup.
The quality picture is mixed too. On the public WGO-Bench, Flora is best at detecting action boundaries but trails Astra overall and falls behind Gemini 3.8 Flash on pure description accuracy. On Nvidia's own DenseCaptionBench, Flora leads across all metrics for human recordings but comes second to Astra for robot videos — and Astra itself evaluates the descriptions in that benchmark while also being one of the models under test. Flora also falsely flags about 11 percent of normal steps as errors, and it sometimes splits longer actions into too many small steps or misdescribes them: in one video, a robot hand wiping a plate with a cloth was described as folding a napkin.
Why cheaper labels matter for robots
Even with those caveats, the release changes the economics of robot training. Labelling large video datasets is expensive and slow, and it is one reason only well-funded labs can train capable humanoid models. A downloadable system handling segmentation, captioning and rewards lets smaller labs and startups build bigger training sets without paying a commercial service such as Scale AI's Physical AI Data Engine — Flora's most direct commercial rival.
It also keeps developers inside Nvidia's ecosystem: Flora runs on Nvidia hardware and feeds its GR00T models. Google's Gemini 4 Argon targets a similar stage of the pipeline. The race over robot training data is increasingly a race over the cheapest complete pipeline — and Nvidia has just given its version away for research.
By Hamza Faisal
