Nvidia’s Flora Brings Cheaper Robot Video Labeling, With Terms To Consider

🔍 Read the full analysis: Nvidia’s Flora Brings Cheaper Robot Video Labeling, With Terms To Consider on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• 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 set of three models that segment, describe and score video of robot tasks, and said its captioning pipeline costs an estimated $0.72 per video hour. That figure is far below the $88.49 per hour Nvidia reports for GPT-6 Astra on its test, but the comparison pairs a modeled GPU cost with measured API usage, and Flora’s license restricts use to research and evaluation.

The three models address different labeling tasks. Flora-TempSeg, with 9 billion parameters, identifies time intervals in which actions occur. The 27-billion-parameter Flora-Captioner writes an English instruction for each interval, while Flora-Rewarder scores a full episode or grades short steps as normal, suboptimal or failures. The models are based on Alibaba’s Qwen3.5 family, according to the paper and model cards.

Nvidia’s cost comparison used 50 episodes totaling 4,483 seconds—about 75 minutes—from its DenseCaptionBench. The company says GPT-6 Astra’s public API charges came to $88.49 per video hour. Flora’s $0.72 figure is a model based on eight GB300 GPUs priced at an assumed $10 per GPU-hour and estimated throughput; Nvidia gives a range of $0.51 to $1.26, excluding storage and other overhead. Dividing the two headline figures produces a ratio of about 122.5, rounded by Nvidia to approximately 125 times.

The benchmark results do not show Flora leading in every category. Across the reported quality comparison, Flora scored 57.1%, compared with 53.7% for GPT-6 Astra, 38.3% for Gemini 3.8 Flash and 30.4% for Claude Opus 5. The paper reports Flora doing better on human videos, action boundaries and some partial-success judgments, while Astra performed better on robot videos, caption accuracy and rubric scoring. Nvidia conducted most of the tests itself.

Lower Costs Could Expand Robot Training Data

Robot-learning systems need more than video: they need descriptions that identify what happened, when it happened and whether an action succeeded. Producing those labels can be costly when applied across large datasets. If Flora’s throughput and cost estimates hold in real deployments, lower-cost labeling could reduce one expense in building and evaluating robot models.

Nvidia says Flora served as a data engine for its GR00T-2 robot model and that it can process up to 200,000 hours of video per day using 512 GB300 GPUs. Those are Nvidia’s reported figures, not an independent production audit. The estimate also concerns captioning; the stated 125-times comparison does not include Flora-Rewarder’s scoring tasks.

For businesses, the terms are a major qualification. The models are released under the NVIDIA OneWay Noncommercial License, which permits research or evaluation use and applies the same limitation to derivatives. Companies therefore cannot treat the release as a free production labeling service under the stated terms. The models may still offer researchers a tool for testing methods and give commercial developers a reference point for costs and performance.

Three Models, Three Labeling Tasks

Robot models are trained on recordings of tasks such as placing an object, passing a tool or folding fabric. The Flora paper describes three types of labels used with that footage: timestamped captions that connect instructions to moments in a clip, process rewards that grade individual steps, and outcome rewards that score an entire attempt. These labels can help researchers organize training data and identify where a robot’s performance went wrong.

The Flora models are available through Hugging Face repositories, with inference code included there, according to the source material. The model cards say they were trained on fewer than 10,000 hours of human and robot video. Testing cited in the paper used an Nvidia H100, while the cost estimate assumes GB300 hardware. The two 27-billion-parameter models also require substantial memory: the source’s estimate is about 55 GB per model for weights in 16-bit precision, placing them beyond typical consumer computers.

“expensive at the scale of modern robot datasets”

— Nvidia, describing the cost of frontier models for robot datasets

Production Costs and Quality Need Validation

The reported 125-times cost gap is not a like-for-like comparison of two measured production bills. GPT-6 Astra’s figure is based on API usage, while Flora’s figure relies on an assumed GPU rental rate and modeled throughput. Nvidia’s paper says the production GPU-hour denominator has not been verified; its estimate also excludes storage and overhead. The test covered about 75 minutes of video selected for Nvidia’s benchmark, so performance and costs on other datasets remain uncertain.

Quality also varies by task and video type. Nvidia’s results show advantages for Flora on some measures, but Astra leads on robot-video captioning accuracy and rubric scoring. The source material does not provide independent replication of the benchmark or establish how the models perform across a wider range of robots and real-world conditions. It also remains unclear whether Nvidia will revise the license or offer a separate commercial-use arrangement.

Independent Tests and Licensing Will Matter

The next useful evidence will be independent evaluations on larger and varied robot-video datasets, alongside measured costs from production workloads that include hardware utilization, storage and operating overhead. Those results would help teams determine whether the modeled savings persist beyond Nvidia’s benchmark and whether quality is adequate for specific labeling jobs.

Organizations considering Flora will also need to review the license before deploying it. Under the published terms, use is limited to research or evaluation, so commercial teams should not assume they can use the models in production or turn modified versions into commercial services. Nvidia’s future licensing guidance, if any, and further benchmark results could change how useful Flora is outside research settings.

Key Questions

What is Nvidia Flora?

Flora is a set of three video models from Nvidia’s GEAR robotics lab. They identify action intervals, describe actions in captions and score robot-task performance.

Is the 72-cent cost a measured production price?

No. Nvidia’s $0.72 per video hour is modeled using an assumed GPU rental price and estimated throughput. The paper says the production GPU-hour denominator has not been verified, and the figure excludes storage and overhead.

Does Flora outperform GPT-6 Astra?

Not across every test. Flora scored higher in the reported overall comparison and performed better on some human-video and partial-success measures, while Astra led on robot videos, caption accuracy and rubric scoring, according to Nvidia’s results.

Can a company use Flora commercially?

The published NVIDIA OneWay Noncommercial License permits research or evaluation use only, and derivatives inherit that restriction. Companies should check the license and seek appropriate terms before any commercial use.

Source: ThorstenMeyerAI.com