🔍 Read the full analysis: The Cost Savings And Licensing Limits To Consider With Nvidia Flora on ThorstenMeyerAI.com
Get audio and creator gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Nvidia’s GEAR lab released Flora, three models for segmenting, captioning and scoring robotics video. Nvidia estimates captioning costs at $0.72 per video hour, compared with $88.49 for GPT-6 Astra on a 50-episode test, but the Flora figure is modeled rather than a measured production bill. The models’ OneWay license limits use to research or evaluation, not commercial production.
Nvidia’s GEAR robotics lab has released Flora, a set of three models that segment, caption and score video of robot tasks, and says its captioning pipeline could cost $0.72 per video hour. That figure is a modeled estimate, not a measured production charge, and the models’ license limits use to research or evaluation rather than commercial deployment.
Flora includes a 9-billion-parameter segmenter that identifies action intervals, a 27-billion-parameter captioner that describes each interval, and a 27-billion-parameter reward model that scores task attempts. Nvidia says the models were built on Alibaba’s Qwen3.5 family and used as a data engine for its GR00T-2 robotics model. The project’s model cards say training used fewer than 10,000 hours of human and robot video.
Nvidia’s cost comparison uses 50 episodes from its DenseCaptionBench, totaling 4,483 seconds of video, or about 75 minutes. It reports GPT-6 Astra API usage at $88.49 per video hour and estimates Flora at $0.72, producing a ratio of about 122.5 to one, rounded by Nvidia to approximately 125 times. Flora’s estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput; it excludes storage and overhead. Nvidia’s paper says the production GPU-hour denominator has not been verified.
The reported quality results are mixed. Across the same benchmark episodes, Flora scored 57.1% and Astra 53.7%, according to Nvidia’s paper. Flora did better on human videos, action boundaries and some cases of partial success, while Astra performed better on robot videos, caption accuracy and rubric scoring. The cost figure applies to captioning; it does not include the reward model. Nvidia ran most of the tests itself, and the provided material does not establish results from independent replication.
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, With Limits
Video labels help robotics teams turn recordings into training and evaluation data. Timestamped descriptions can identify which action occurs when; step-level scores can flag failures or suboptimal movements; and whole-episode scores can help track task outcomes. If a labeling pipeline can be run at substantially lower cost, teams may be able to process more video within a fixed budget, reducing one expense in building robot systems.
That prospect needs to be weighed against the estimate’s assumptions and the license. The $0.72 figure is not an invoice or production measurement, and Nvidia’s test covers a limited benchmark set. Companies cannot treat Flora as a free commercial labeling service under the stated license. For business users, the release is better understood as a research resource and a reference point for costs and performance, subject to legal review of the applicable terms.
Lower annotation costs would not by themselves solve other constraints on robotics development. Teams still need suitable video, physical robots and compute, and they must address safety and reliability. The reported benchmarks also show that a lower-cost system does not lead on every type of video or scoring task.
NVIDIA RTX workstation GPU for AI model inference
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Three Models for Robot Video
Robotics models learn from recordings of actions such as picking up objects or folding fabric. The Flora paper describes three related labeling needs: timestamped captions tied to video intervals, rewards for individual steps, and scores for complete attempts. Flora-TempSeg identifies intervals, Flora-Captioner writes an instruction for each, and Flora-Rewarder can grade short steps or score an episode against a rubric.
The captioner can take video from one camera or combine a main camera with wrist cameras, according to the model description. The models are available through Hugging Face repositories, with inference code included there. The stated test setup used an Nvidia H100 on Linux. The source material estimates that each 27-billion-parameter model needs roughly 55 GB for weights at 16-bit precision, suggesting that local use calls for data-center-class hardware rather than an ordinary gaming computer.
Nvidia’s headline comparison is narrower than the full Flora suite: it compares captioning costs on a selected set of episodes. The benchmark includes human hands, two-finger grippers and dexterous robot hands. The paper also reports results on other tests, including WGO-Bench and RoboRewardBench, where Flora performs better on some measures. Those separate results do not turn the modeled captioning estimate into a measured end-to-end operating cost.
“research or evaluation purposes only”
— Nvidia, describing Flora’s license
Cost and License Questions
The cost comparison remains an estimate rather than a verified production result. Nvidia modeled Flora’s GPU throughput and assumed a rental rate of $10 per GPU-hour; its paper gives scenarios from $0.51 to $1.26 per video hour and says storage and overhead are excluded. The Astra figure, by contrast, is based on reported API usage for the benchmark, though the source says 106 of 7,062 attempts had no usage reported, making the total a known-usage subtotal.
The test contains about 75 minutes of video selected from Nvidia’s own benchmark, so it does not establish what costs or quality would look like across larger, different or independently chosen datasets. Results also vary by task: Flora’s aggregate score was higher in this comparison, but Astra led on several reported measures. The source material does not provide independent validation of the headline cost ratio.
The license’s practical application to particular business arrangements, derivatives or evaluation workflows may require legal interpretation. The stated terms restrict the models to research or evaluation, and derivatives inherit that limit; the source also says Nvidia may use the models commercially. The available material does not detail how every form of internal corporate testing would be treated.
Independent Tests and License Review
The next useful evidence would be independent tests using broader robotics datasets and published, measured GPU utilization, including storage and other operating costs. Such work could show whether Flora’s estimated advantage holds under different hardware prices, workloads and video types, and whether its quality trade-offs are acceptable for particular tasks.
Organizations considering the models should check the OneWay Noncommercial License and assess their intended use before deployment. Nvidia has released the models and supporting paper, but the source material does not state a date for a commercial license, a production cost audit or an independent benchmark. Until those details are available, the reported savings should be treated as a modeled comparison, not a confirmed operating cost for commercial robotics systems.
Key Questions
What is Nvidia Flora?
Flora is a three-model suite for labeling robotics video: one model finds action intervals, one captions them, and one scores steps or complete attempts.
Is Flora confirmed to cost 72 cents per video hour?
No. Nvidia reports $0.72 per video hour as a modeled estimate based on assumed GPU rental costs and throughput. The paper says the production GPU-hour denominator has not been verified, and the estimate excludes storage and overhead.
How does Nvidia calculate the approximately 125-times comparison?
Nvidia compares its modeled Flora estimate of $0.72 per video hour with $88.49 per hour based on GPT-6 Astra API usage on 50 benchmark episodes. Dividing those figures gives about 122.5, which Nvidia rounds to approximately 125 times.
Can companies use Flora commercially?
The models are distributed under the NVIDIA OneWay Noncommercial License, which states that use is for research or evaluation purposes only. The source material says derivatives inherit that limit; companies should review the license for their specific plans.
Does Flora outperform GPT-6 Astra on every test?
No. Flora scored higher on the reported overall benchmark comparison and performed better on some human-video, action-boundary and partial-success measures. Astra led on robot videos, caption accuracy and rubric scoring, according to Nvidia’s paper.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
