Nvidia Flora’s Robot Video Labeling: Where The Cost Savings Meet The License

🔍 Read the full analysis: Nvidia Flora’s Robot Video Labeling: Where The Cost Savings Meet The License on ThorstenMeyerAI.com

Nvidia’s GEAR robotics lab has released Flora, a set of three models that identify actions in video, describe them and score robot performance. Nvidia estimates that Flora’s captioning can cost $0.72 per hour of video, versus $88.49 for GPT-6 Astra in a test described in the paper. The comparison is not a measured, like-for-like production bill: Flora’s figure is modeled, and its license permits research and evaluation, not commercial use.

Flora consists of three models built on Alibaba’s Qwen3.5. Flora-TempSeg, with 9 billion parameters, identifies the time intervals when actions occur. The 27-billion-parameter Flora-Captioner writes an instruction for each interval, while the 27-billion-parameter Flora-Rewarder scores an attempt overall, grades steps, or evaluates performance against a user’s rubric. Nvidia says the models were used as a data engine for its GR00T-2 robot models.

The cost comparison comes from 50 episodes totaling 4,483 seconds—about 75 minutes—selected for Nvidia’s DenseCaptionBench. GPT-6 Astra’s $88.49 per video hour is based on reported API usage at public rates from October 4, 2026. Flora’s $0.72 estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput; the paper gives a range of $0.51 to $1.26 and excludes storage and other overhead. The comparison covers captioning, not the reward model.

On that same test, the paper reports a 57.1% score for Flora and 53.7% for Astra. Results elsewhere are mixed: Flora performs better on the human-video subset and on some action-boundary and partial-success evaluations, while Astra leads on robot-video caption accuracy and rubric scoring, according to the paper. Nvidia conducted most of the evaluations itself, and the reported benchmark results should not be read as an independent assessment of real-world performance.

Lower Labeling Costs, Narrower Access

Labeling footage is part of preparing the data used to train and assess robot models. Tools that locate actions, describe them and identify unsuccessful steps could reduce the labor or computing expense of organizing large video collections. Nvidia says Flora can process up to 200,000 hours of video per day on 512 GB300 GPUs; that is a company-reported capacity, not a demonstrated general service level.

The license changes what the cost claim means for companies. Flora’s models are released under the NVIDIA OneWay Noncommercial License, which limits use to research or evaluation and applies the same restriction to derivatives. A company may use the release to study the approach or compare results, but the supplied terms do not make it a free commercial labeling pipeline. The gap between a low modeled inference cost and the expense of a usable production system also includes hardware, operations, data handling and validation.

For robotics researchers, Flora offers models and a benchmarkable alternative to general-purpose services. For businesses, the results may serve as a cost and performance reference, but commercial deployment would require rights beyond those stated in this release. Cheaper labeling may ease one data-processing burden; it does not by itself resolve the costs of collecting robot footage, building hardware or establishing safe performance.

Three Models for Video Labels

Robot-learning data can include recordings from robot cameras and from people performing tasks. The Flora paper describes several kinds of labels needed for training and evaluation: timestamped action captions, step-by-step process rewards and scores for an overall outcome. These labels help researchers locate actions and compare successful, incomplete or failed attempts. Flora separates those functions across its three models rather than using a single model for every task.

The model cards say Flora was trained on under 10,000 hours of human and robot video. Nvidia reports testing the inference code on Linux with an H100. The source estimates that each 27-billion-parameter model may require roughly 55 GB for weights in 16-bit precision, placing deployment in data-center-class hardware rather than typical consumer computers. Those compute requirements and the noncommercial terms matter alongside the per-video-hour estimate.

Nvidia’s paper documents the cost calculation in an appendix. It compares a modeled GPU budget for Flora with API usage recorded for Astra, and says some Astra attempts reported no usage. The paper describes the Flora production GPU-hour denominator as unverified. The resulting approximately 125-times figure is the rounded ratio of $88.49 to $0.72, not an independently measured saving from operating both systems in production.

„research or evaluation purposes only“

— NVIDIA OneWay Noncommercial License

Unverified Costs and License Scope

Flora’s production cost has not been established by the reported comparison. The estimate depends on assumed GPU pricing and throughput, excludes storage and overhead, and uses a denominator the paper says has not been verified. The test also covers about 75 minutes of selected video, so it does not show how costs or quality would scale across a company’s full dataset.

The benchmark results come largely from Nvidia’s own evaluation. The available material does not establish how Flora would perform across other tasks, camera setups, data distributions or independent tests. Nor does the stated noncommercial license describe a general commercial permission. Businesses would need to review the license and obtain appropriate rights before relying on the models in a commercial workflow.

Independent Tests and Commercial Terms

The next useful checks are independent evaluations on larger and more varied robot-video sets, together with published production measurements that include GPU utilization, storage and operating overhead. Those results would help establish whether the modeled cost advantage persists outside the episodes and hardware assumptions in Nvidia’s paper.

For prospective users, the immediate steps are to examine the model cards, license and inference requirements, then distinguish research evaluation from commercial deployment. Nvidia’s release does not, in the supplied material, announce a commercial license or a production service. Whether either becomes available—and how Flora performs in external tests—remains open.

Key Questions

What does Nvidia Flora do?

Flora is a set of three models: one segments video into action intervals, one writes captions for those intervals, and one scores robot attempts or individual steps.

Is the $0.72-per-hour figure a measured operating cost?

No. Nvidia’s figure is a modeled estimate based on assumed GPU rental costs and throughput. The paper gives a range of $0.51 to $1.26 per video hour and says the production GPU-hour denominator has not been verified.

Does Flora outperform GPT-6 Astra?

Not across every test. Flora scored 57.1% versus Astra’s 53.7% on the reported 50-episode comparison and did better on some human-video and partial-success evaluations. The paper reports Astra ahead on robot-video caption accuracy and rubric scoring.

Can a company use Flora commercially?

The stated NVIDIA OneWay Noncommercial License allows use for research or evaluation only, and derivatives inherit that restriction. The supplied material does not provide commercial-use rights.

What is not included in Nvidia’s cost comparison?

The comparison covers captioning and excludes storage and overhead. It does not include Flora’s reward model in the 125-times calculation or establish total production costs for a commercial system.

Source: ThorstenMeyerAI.com