Nvidia Open-Sources Flora, a Robot Video Labeler That Costs 72 Cents an Hour Instead of 88 Dollars

Nvidia researchers published Flora on October 25, an open model suite that turns raw robot video into the three kinds of labels a robotics foundation model needs. The work comes out of Nvidia GEAR, the company research lab that also builds the GR00T robot models, and the weights are on Hugging Face under a collection named for the project.

The design answers a practical split. Timestamped captions are needed to filter and balance pretraining data, process rewards are needed during reinforcement-learning post-training, and outcome rewards are needed to monitor a robot after deployment. Most annotation tooling covers one of the three. General models can produce all three, but the volume of video involved makes per-call pricing unworkable, which is the gap Flora targets.

Flora is three fine-tuned Qwen3.5 models doing different jobs. A 9B model cuts video into action segments, a 27B model writes descriptions that account for what happened before and after each action, and a second 27B model called Flora-Rewarder scores episodes from 1 to 5, classifies every half-second step as normal, suboptimal or failure, and grades against self-defined rubrics. Segmentation and captioning produce timestamped text; one reward model supplies all three rating types.

The headline number is cost. Nvidia puts one hour of labeled video at about 72 cents with Flora, against roughly 88.50 dollars with GPT-6 Astra and about 42 dollars with Claude Opus 5 in the same test. That is the source of the 125x claim. The figure is a projection, not a metered bill: Nvidia calculated it from an assumed rental price of 10 dollars per GPU per hour. On a 50-video internal test the company reports 57.1 percent annotation quality for Flora at that price point, against 53.7 percent for Astra at the higher one, with Astra allowed its full reasoning budget.

Public benchmarks are more mixed, and the paper says so. On WGO-Bench, Flora leads segmentation at 50.3 F1@0.75 against 46.0 for Astra, but trails badly on gold-window accuracy at 64.7 versus 77.4, and Gemini 3.8 Flash is more accurate on pure description. On Nvidia's own DenseCaptionBench, Flora leads every metric for human-recorded video, with joint point F1 of 37.4 against 29.7 for Astra, but falls to second on robot video at 51.0 against 57.9. The researchers also concede a methodological wrinkle: GPT-6 Astra is one of the models being tested and also acts as the judge of description quality.

Where the gap is widest is error detection inside an episode. Flora-Rewarder catches 53.3 percent of failed steps, while no competitor clears 11.7 percent and Astra catches none at all, according to the paper, which attributes the difference to partially completed tasks. When a robot tips a bottle over instead of knocking it off a table, Flora gives 3 out of 5 to match the human reference and Astra gives 1. The cost of that sensitivity is over-flagging: Flora marks about 11 percent of normal steps as errors. In a clip dense with short hand movements it identified 34 of 37 reference actions where Astra found four, and in another example it described a robot hand wiping a plate with a cloth as folding a napkin.

Scale is the part that separates Flora from a research demo. Nvidia says the engine runs during training of its GR00T models and annotates up to 200,000 hours of video per day across 512 GB300 GPUs, roughly 390 hours per GPU per day. The company frames the release as an open alternative to commercial annotation services such as Scale AI's Physical AI Data Engine, and as another way to keep robotics developers inside its hardware ecosystem.

The competitive read is straightforward. Google's Gemini 4 Argon was pitched at a similar slice of the pipeline, identifying work steps in training video and flagging failed actions, so cheap reward labeling is becoming table stakes rather than a differentiator. The number that matters for anyone adopting Flora is not the 125x. It is the 11 percent false-flag rate on normal steps, because a reward model that cries wolf a tenth of the time will teach a policy to distrust its own successes during post-training. Teams should validate Flora-Rewarder on their own failure taxonomy before pointing it at a large corpus.