🔍 Read the full analysis: Flora’s Robot Video Labeling Savings Sound Big; The License Adds A Catch on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Nvidia’s GEAR lab released Flora, three models for labeling robot videos, and reports a modeled cost of $0.72 per video hour compared with $88.49 for GPT-6 Astra on a selected test set. The comparison is not a measured production-cost test, performance varies across benchmarks, and Flora’s license limits use to research or evaluation rather than commercial deployment.
Nvidia’s GEAR robotics lab has released Flora, a set of three models designed to label video used in robot training, and says the system’s estimated labeling cost is $0.72 per hour of video, compared with $88.49 for GPT-6 Astra in a test reported by Nvidia. The headline cost gap is based on modeled GPU expenses rather than a measured production bill, and the models’ OneWay Noncommercial License restricts use to research or evaluation.
Flora is made up of three models: a 9-billion-parameter temporal segmenter that identifies when actions begin and end; a 27-billion-parameter captioner that describes each interval; and a 27-billion-parameter rewarder that scores full attempts or grades steps as normal, suboptimal or failures. Nvidia says the models were built on Alibaba’s Qwen3.5 models and used as a data engine for its GR00T-2 robot models.
The cost comparison uses 50 episodes from Nvidia’s DenseCaptionBench, totaling 4,483 seconds of video, or about 75 minutes. Nvidia reports GPT-6 Astra API costs of $88.49 per video hour and estimates Flora at $0.72, producing a ratio of 122.5 to one, rounded in the paper to approximately 125 times. Flora’s estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and excludes storage and other overhead. Nvidia’s paper says the production GPU-hour denominator has not been verified; its modeled scenarios range from $0.51 to $1.26 per video hour.
On the same episodes, the paper reports a 57.1% score for Flora and 53.7% for Astra. The results vary by task: Flora performs better on human-video examples, action boundaries and some partial-success cases, while Astra leads on robot videos, caption accuracy and rubric scoring, according to the benchmark results. The reported 125-times cost comparison concerns captioning, not the separate reward model.
Infographic 01 · What Flora does
Three models that turn robot video into training labels.
Flora cuts a recording into actions, describes each one and grades how well it went. The 200,000-hour figure is Nvidia’s own production claim.
Flora-TempSeg
9B · adapter on Qwen3.5-9B
- Job: Finds where each action starts and ends
- In: Video windows (one camera, or main + two wrist cameras)
- Out: Time intervals as JSON
Flora-Captioner
27B · fine-tuned Qwen3.5-27B
- Job: Describes each action, using the video before and after
- In: One interval plus context
- Out: One English instruction, e.g. “Place the blue mug on the tray with your right hand.”
Flora-Rewarder
27B · fine-tuned Qwen3.5-27B
- Job: Judges how well the robot did
- In: An episode or a 16-second window, plus the task
- Out: Score 1–5 per episode · normal / suboptimal / failure per 0.5 s · rubric scores
Three labels, three uses
Why robot models need each one, per the paper
| Label | Used for | Made by |
|---|---|---|
| Timestamped captions | Filtering and rebalancing pretraining data | TempSeg + Captioner |
| Process rewards | Reinforcement learning after pretraining: which step went wrong | Rewarder |
| Outcome rewards | Monitoring robots after deployment | Rewarder |
What it is not
model cards
- Not a robot controller: outputs are annotations only
- Not meant to be the sole safety monitor for a physical system
- Not for commercial or production use (noncommercial license)
- Hardware: tested on Nvidia H100, Linux; ~55 GB of weights per 27B model (estimate)
Sources: Flora paper, Nvidia (20 Oct 2026); Hugging Face model cards (15 Oct 2026); Nvidia GEAR project page. Weight-size estimate: Thorsten Meyer (27B parameters × 2 bytes).
Powered by Thorsten Meyer AI
Lower Labeling Costs, Limited Deployment
Video labels help robotics teams connect what a camera records with what a robot should do, identify where an attempt went wrong and evaluate performance. If the cost estimates hold at larger scale, cheaper labeling could reduce one expense in preparing large robot-training datasets. Nvidia says Flora can process up to 200,000 hours of video per day using 512 GB300 GPUs; that capacity claim is Nvidia’s, not an independently verified production result in the supplied material.
The license changes the practical value for companies. The noncommercial terms allow research or evaluation, while derivatives inherit that restriction. Businesses therefore cannot treat the published models as a free commercial labeling service under these terms. The release may still offer a benchmark and technical reference, but commercial users would need different licensing or permission. Lower labeling costs alone also do not resolve the expense of collecting robot data, building hardware or establishing safe operation.
NVIDIA GPU workstation for AI model development
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Three Models for Robot Video
Robot-learning datasets include footage of people and machines performing tasks, such as placing objects or folding fabric. According to the Flora paper, those recordings need several kinds of annotation: timestamped descriptions of actions, step-level feedback about success or failure, and scores for an entire attempt. These labels can help teams organize training data and assess robot behavior.
The GEAR lab’s release packages the work into separate models for segmentation, captions and rewards. The paper says Flora was trained on fewer than 10,000 hours of human and robot video. Nvidia’s cost comparison is narrower than the full labeling task: it compares captioning costs on a fixed set of episodes and pits a modeled GPU budget against measured API usage. The paper also reports that 106 of 7,062 Astra attempts had no usage data, making its API total a known-usage subtotal.
Production Costs and Commercial Rights
It remains unclear whether Flora will reach the estimated cost at production scale. The $0.72 figure assumes a rental rate and modeled throughput; it does not include storage or overhead, and Nvidia says the production GPU-hour denominator has not been verified. The comparison also uses about 75 minutes of selected test video, so it does not establish costs across other datasets, workflows or operating conditions.
The benchmark results are not uniform, and most tests were run by Nvidia, according to the source material. The reported score advantage on the combined test does not mean Flora leads on every category. The available material does not establish whether Nvidia will offer a commercial license, whether outside users can reproduce the throughput estimate, or how the models perform across a wider range of robots and real-world tasks.
Reproduction and Licensing Will Matter
The next useful evidence would be independent tests at production scale that report throughput, complete GPU and infrastructure costs, and results across more video datasets. That would help determine whether the modeled savings persist beyond Nvidia’s fixed benchmark and assumed rental price.
For companies weighing deployment, the immediate next step is to review the license and seek clarification or separate permission before using Flora commercially. Nvidia has not, in the supplied material, announced a commercial-use option or a timetable for further validation. Until those details emerge, the cost claim should be read as a promising estimate, not a confirmed commercial operating cost.
Key Questions
What is Nvidia Flora?
Flora is a three-model system from Nvidia’s GEAR robotics lab for segmenting robot video, writing captions for actions and scoring robot attempts or steps.
Is Flora really 125 times cheaper than GPT-6 Astra?
Nvidia reports a ratio of about 122.5 to one on its test, rounded to approximately 125 times. Flora’s $0.72-per-video-hour cost is modeled using assumed GPU rental costs, while Astra’s $88.49 figure comes from measured API usage; the test covered about 75 minutes of selected video.
Can businesses use Flora commercially?
The stated NVIDIA OneWay Noncommercial License permits research or evaluation only, and derivatives inherit the restriction. The supplied material does not describe a commercial-use license.
Does Flora outperform GPT-6 Astra on every task?
No. The reported results are mixed: Flora leads on some human-video, action-boundary and partial-success measures, while Astra leads on robot videos, caption accuracy and rubric scoring.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
