The New Cost Equation For Robot Video Labeling: Nvidia Flora And Its License

🔍 Read the full analysis: The New Cost Equation For Robot Video Labeling: Nvidia Flora And Its License 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 has released Flora, a set of three models for labeling video used in robot training, and estimates that its captioning pipeline costs $0.72 per hour of video compared with $88.49 for GPT-6 Astra on the company’s test set. The cost comparison is not like-for-like: Flora’s figure is a modeled GPU budget, while Astra’s is based on measured API usage, and Nvidia’s license permits research or evaluation only, not commercial deployment.

Flora includes a 9-billion-parameter temporal segmenter that identifies when actions begin and end, a 27-billion-parameter captioner that writes instructions for those segments, and a 27-billion-parameter reward model that scores a robot’s performance. The models are built on Alibaba’s Qwen3.5 family. Nvidia says Flora was used as a data engine for its GR00T-2 robot model and was trained on fewer than 10,000 hours of human and robot video.

The paper’s cost test used 50 episodes totaling 4,483 seconds, or about 75 minutes, from Nvidia’s DenseCaptionBench. Nvidia reports $88.49 per video hour for GPT-6 Astra at maximum reasoning effort using public API rates from October 4, 2026. Flora’s $0.72 estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput; Nvidia gives a range of $0.51 to $1.26, excluding storage and other overhead. The approximately 125-fold comparison is the result of dividing $88.49 by $0.72.

On the same test, Nvidia reports Flora scored 57.1% and Astra 53.7%, with Gemini 3.8 Flash at 38.3% and Claude Opus 5 at 30.4%. Those figures do not establish that Flora is better across tasks: the paper’s other benchmark results are mixed, and the 125-fold cost claim covers captioning, not Flora’s reward model. Nvidia also says the GPU-hour denominator for production has not been verified.

Lower Labeling Costs, With Limits

Robot developers need video labels to identify actions, describe what happened and assess whether a robot succeeded. If labeling is expensive, it can constrain how much recorded data a team can process. A substantially cheaper labeling option could reduce one cost in building and testing robot systems, particularly when datasets contain large volumes of video.

But the headline figure is not a verified production price. It compares a modeled GPU cost with measured API usage on a relatively small test set selected by Nvidia, and the estimates omit storage and overhead. Results also vary by task: Flora performs well on some human-video, action-boundary and partial-success tests, while Astra leads on some robot-video and scoring measures. The release is best read as evidence of a potential cost reduction, not proof that companies can replace paid services at the stated rate.

The license is a separate practical constraint. The models are distributed under the NVIDIA OneWay Noncommercial License, which limits use to research or evaluation and applies the same restriction to derivatives. That makes Flora potentially useful for researchers comparing methods or testing workflows, but not a free commercial labeling service under the stated terms. Nvidia’s own commercial use of its models does not grant that right to other users.

Three Models for Robot Video

Robot-learning systems use recorded demonstrations to connect actions with outcomes. A video might show a gripper placing a mug, a person folding a towel or a robot failing to move an object. To make those recordings useful for training, developers may need timestamps and captions, grades for individual steps, and scores for an entire attempt.

Flora divides those jobs among three models. Flora-TempSeg identifies action intervals; Flora-Captioner describes them in English, using surrounding video for context; and Flora-Rewarder can assign an overall score from one to five, label half-second segments as normal, suboptimal or failure, or assess performance against a user-provided rubric. The captioner can work with one camera or a main camera plus wrist cameras, according to the model cards.

The models and inference code are hosted in Hugging Face repositories, and Nvidia’s paper describes testing on an H100 system. The source material estimates that each 27-billion-parameter model would require roughly 55 GB of memory for weights in 16-bit precision, indicating that local use needs data-center-class hardware rather than a typical consumer PC.

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

— Nvidia, in the Flora paper

Production Costs and License Scope

The reported cost advantage remains an estimate rather than a production measurement. The source does not establish how the figure would change with different hardware rental rates, video mixes, processing requirements or operational overhead. Nvidia’s test covers 50 episodes, and the company says the production GPU-hour denominator has not been verified. The reported Astra usage total also excludes 106 of 7,062 attempts for which no usage was reported.

Benchmark results do not settle how Flora would perform across a company’s own data. The provided results show wins and losses across human and robot videos, captioning, action segmentation and reward scoring. The available material also does not provide an independent replication of Nvidia’s cost or quality results. It remains unclear how the models would compare at larger scale and under varied deployment conditions.

For commercial users, the stated license rules out treating the release as a production-ready, free-to-use pipeline without separate authorization or different terms. The source material does not describe any commercial license option or indicate whether Nvidia plans to offer one.

Testing Beyond Nvidia’s Benchmark

The next useful evidence will come from independent tests that publish their video sets, hardware, throughput, quality measures and full operating costs. Those results could show whether Flora’s modeled savings persist outside Nvidia’s selected episodes and whether its weaker results on some tasks affect practical use.

Teams considering Flora will also need to check the license against their intended use before adopting it. Researchers can evaluate the models under the stated terms, while commercial deployment requires rights not granted by the noncommercial license. Nvidia has not provided, in the supplied material, a schedule for further releases, a commercial licensing path or updated production-cost measurements.

Key Questions

What is Nvidia Flora?

Flora is a set of three video models for robot-learning data: one segments actions in time, one captions them, and one scores robot performance.

Is the $0.72-per-hour cost a measured production price?

No. Nvidia describes it as a modeled estimate based on eight GB300 GPUs at an assumed rental price of $10 per GPU-hour and modeled throughput. Storage and other overhead are excluded.

How was the roughly 125-fold comparison calculated?

Nvidia divided the reported $88.49 per video hour for GPT-6 Astra by Flora’s modeled $0.72 per hour. The test used 50 episodes, about 75 minutes of video; the two figures come from different cost methods.

Can a company use Flora commercially?

The stated NVIDIA OneWay Noncommercial License allows use for research or evaluation only and extends that restriction to derivatives. The supplied material does not identify a commercial-use option.

Does Flora outperform GPT-6 Astra on every test?

No. Nvidia reports Flora ahead on its DenseCaptionBench aggregate and some other tasks, but Astra leads on certain robot-video, caption-accuracy and rubric-scoring results. Performance depends on the benchmark and task.

Source: ThorstenMeyerAI.com