Nvidia Flora Lowers Robot Video Labeling Costs; Its License Deserves Attention

🔍 Read the full analysis: Nvidia Flora Lowers Robot Video Labeling Costs; Its License Deserves Attention on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Nvidia’s GEAR robotics lab released Flora, a three-model suite that segments, captions and scores video of robot tasks, and says its captioning can cost about $0.72 per hour of video. That cost is a model based on assumed GPU rental rates and throughput—not a measured production bill—and the models’ license permits research or evaluation, not commercial use.

Flora includes a 9-billion-parameter temporal segmenter that identifies where actions begin and end, a 27-billion-parameter captioner that describes each segment, and a 27-billion-parameter reward model that grades a task or individual steps. Nvidia says the tools were used as a data engine for its GR00T-2 robot models. The models are built on Alibaba’s Qwen3.5 family and are available through Hugging Face repositories, according to the source material.

The cost comparison in Nvidia’s paper uses 50 episodes totaling 4,483 seconds, or about 75 minutes, from Nvidia’s DenseCaptionBench. Nvidia reports an API cost of $88.49 per video hour for GPT-6 Astra and estimates Flora at $0.72 per hour. The estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput; the paper says the production GPU-hour denominator has not been verified. The stated ratio is about 122.5 to one, rounded by Nvidia to approximately 125 times.

The comparison concerns captioning, not all three Flora models. Nvidia’s reported quality score on the same test episodes was 57.1% for Flora and 53.7% for Astra. Results were not uniformly in Flora’s favor: the paper reports Flora doing better on human videos, action boundaries and partial-success judgments, while Astra performed better on robot videos, caption accuracy and rubric scoring. Nvidia conducted most of the tests described in the source material.

Lower Costs, With Trade-Offs

Video labeling is a substantial part of building robot-learning datasets: systems need to identify actions, describe what happened and assess whether a step succeeded. If the cost difference holds at larger scale, less expensive labeling could reduce one data-processing bottleneck for robotics research and model development. Nvidia says Flora can process up to 200,000 hours of video per day using 512 GB300 GPUs, but that is a company-reported capacity figure, not an independently verified production result.

The license sets a clear limit for businesses. All three models are provided under the NVIDIA OneWay Noncommercial License, which allows research or evaluation and passes the restriction to derivatives, according to the source material. That means the models may be useful for testing workflows and comparing approaches, but organizations cannot treat them as a free commercial labeling service under the stated terms. The source says Nvidia may use the models commercially.

The broader effect should not be overstated. Lower labeling costs do not by themselves solve the expense of collecting representative robot data, buying and operating hardware, or making robots safe in real environments. Flora’s results also vary across video types and scoring tasks. For readers evaluating the headline cost claim, the key distinction is that Astra’s figure is based on API usage, while Flora’s is a modeled GPU budget.

Three Models, Three Labeling Tasks

Robot-learning systems use video from robots and people performing tasks—for example, moving an object or folding fabric. The source material describes three types of labels used in this work: timestamped captions that tie instructions to parts of a video, step-level rewards that identify normal, suboptimal or failed actions, and episode-level scores that summarize an attempt. Flora separates those functions across its segmenter, captioner and rewarder.

Nvidia’s paper says Flora was trained on fewer than 10,000 hours of robot and human video. The evaluation set included 20 human-hand episodes, 15 two-finger-gripper episodes and 15 dexterous-robot-hand episodes. The source material also reports that the 27-billion-parameter models require substantial memory in 16-bit precision, putting local use closer to data-center hardware than an ordinary gaming PC. Those details matter when weighing a low estimated per-video cost against the infrastructure needed to run the models.

The comparison’s scope is limited. Its 75 minutes of test footage came from a benchmark created by Nvidia, and the cost calculation covers the captioning task. The article’s supplied source notes that 106 of 7,062 Astra attempts reported no usage, making the $88.49-per-hour figure a known-usage subtotal. It also gives Flora cost scenarios ranging from $0.51 to $1.26 per video hour, excluding storage and other overhead.

“The production GPU-hour denominator has not been verified.”

— Nvidia, in the Flora paper

Cost and Benchmark Limits

The Flora cost is not an invoice from a production deployment. It relies on assumed GPU rental prices and modeled throughput, and the paper says its production GPU-hour denominator has not been verified. The comparison also excludes storage and overhead, while Astra’s API figure is based on recorded usage with 106 attempts reporting no usage. The figures therefore do not establish a like-for-like production cost across all operating conditions.

It is also unclear from the supplied material how Flora would perform across a broader range of independently selected video datasets, camera setups and real-world robot tasks. The reported quality results are mixed, and the test set was Nvidia’s own benchmark. The 125-times comparison applies to captioning only; it should not be read as a cost comparison for the reward model or the complete labeling workflow. Commercial use is restricted by the stated license, though the supplied material does not detail every possible licensing scenario.

What Deployment Would Require

The immediate next step for researchers and robotics teams is to test the models on their own video and hardware, while checking the OneWay Noncommercial License before adopting them. More evidence would be needed to establish real operating costs: verified throughput, GPU utilization, storage and engineering expenses, and performance across datasets not selected by Nvidia.

For commercial deployment, organizations will need to resolve the license restriction rather than assume the models can be used in a paid production pipeline. Nvidia’s release makes Flora available for research and evaluation, but the supplied information does not announce a commercial license or a measured production-cost study. Until those details change, the cost advantage remains an Nvidia estimate, and the commercial pathway remains unsettled.

Key Questions

What is Nvidia Flora?

Flora is a suite of three models for robot-video labeling: one identifies action segments, one writes captions, and one scores task outcomes or steps.

Is the $0.72-per-hour cost measured?

No. Nvidia’s $0.72 figure is modeled using assumed GPU rental costs and throughput. Its paper says the production GPU-hour denominator has not been verified.

Does Flora beat GPT-6 Astra on every test?

No. Nvidia reports Flora scoring higher overall on the cited test set and doing better on some human-video and partial-success tasks. Astra performed better on some robot-video, caption-accuracy and rubric-scoring results.

Can a company use Flora commercially?

The stated NVIDIA OneWay Noncommercial License permits research or evaluation and restricts commercial use. The source material does not describe a separate commercial license.

Does the 125-times figure cover all of Flora?

No. The comparison is for captioning on a 50-episode test set. It does not include the cost of the reward model or establish the full cost of a production labeling workflow.

Source: ThorstenMeyerAI.com