🔍 Read the full analysis: Flora Could Cut Robot Video Labeling Costs By 125×, But Terms Matter on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Nvidia’s GEAR lab released Flora, a set of three models for segmenting, captioning and scoring robot videos. Nvidia estimates captioning costs about $0.72 per video hour, compared with $88.49 for GPT-6 Astra on a 50-episode test, but Flora’s cost is modeled, its quality results are mixed, and its license permits research and evaluation rather than commercial use.
Nvidia’s GEAR robotics lab has released Flora, three models designed to label video used in robot training, and estimates their captioning cost at $0.72 per video hour—about 122.5 times below the $88.49-per-hour GPT-6 Astra figure in its comparison. The cost gap is Nvidia’s estimate, not a measured production bill, and Flora’s license allows research and evaluation only, limiting its direct use by commercial robotics companies.
Flora combines three models built on Alibaba’s Qwen3.5 family. Flora-TempSeg, with 9 billion parameters, identifies the time intervals in which actions occur. The 27-billion-parameter Flora-Captioner writes a description for each interval, while the 27-billion-parameter Flora-Rewarder scores an episode or grades short steps as normal, suboptimal or failed. Nvidia says Flora served as a data engine for its GR00T-2 robot model.
The paper’s cost comparison used 50 episodes from Nvidia’s DenseCaptionBench, totaling 4,483 seconds of video, or about 75 minutes. Nvidia reports $88.49 per video hour for GPT-6 Astra using public API rates and $0.72 for Flora. The latter is modeled using eight GB300 GPUs at an assumed $10 per GPU-hour, with estimated throughput; Nvidia’s paper says the production GPU-hour denominator has not been verified. The estimate excludes storage and overhead and gives scenarios from $0.51 to $1.26 per video hour. The comparison covers captioning, not Flora’s reward model.
On the same episodes, Nvidia reports an overall score of 57.1% for Flora and 53.7% for Astra. The results vary by task: Flora scored better on human videos, action boundaries and partial-success cases, while Astra led on robot videos, caption accuracy and rubric scoring. Nvidia conducted most of the tests, so the figures are not an independent evaluation. The source account also notes that 106 of 7,062 Astra attempts reported no usage; the $88.49 figure is a known-usage subtotal.
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
The Cost of Robot Training Data
Labeling video is one step in preparing the large datasets used to train robot models. Timestamped captions can help researchers locate and balance examples; step-level rewards can identify where a robot failed; episode-level scores can track outcomes. If a lower-cost system can produce useful labels at scale, it could reduce one expense in building and evaluating robot models.
But the advertised ratio should not be read as a settled operating-cost comparison. It pairs a measured API usage figure with a modeled GPU budget, using a limited set of episodes selected by Nvidia. Actual costs could change with hardware rental rates, utilization, throughput, storage and engineering overhead. And cheaper output is not automatically interchangeable with stronger labels: the benchmark shows Flora ahead on some tasks and behind on others.
The license is a separate constraint for companies weighing adoption. The models are distributed under the NVIDIA OneWay Noncommercial License, which permits research or evaluation and restricts commercial use, including derivatives. That makes Flora potentially useful for testing or as a reference point, but the published terms do not provide a straightforward free production pipeline for a business. The source material says Nvidia may use the models commercially.
Products Worth Considering
Dell XPS 13 9350 Laptop, 13.4 FHD+ Display, Intel Ultra 7 256V, 16GB DDR5, 1TB SSD
Dell UltraSharp 32 U3226QER 4K Thunderbolt Hub Monitor
Canon IVY 2 Mini Photo Printer Bundle, Pure White
As an affiliate, we earn on qualifying purchases.
Three Models, Three Labeling Tasks
Robot-learning systems use video of robots and people carrying out tasks, such as placing a mug or folding a towel. According to the Flora paper, model developers may need labels that mark when an action occurs, describe what happened, and score whether a step or full attempt succeeded. These labels help organize training data and assess robot performance.
Flora packages those jobs in separate models rather than one general-purpose system. Its captioner can use video context before and after an interval; it supports one camera or a main camera paired with wrist cameras. The rewarder can score an episode on a five-point scale, grade steps, or assess performance against a user-provided rubric. The models are available through Hugging Face repositories with inference code, according to the source material.
Nvidia says Flora was trained on fewer than 10,000 hours of robot and human video. The source account estimates that each 27-billion-parameter model requires roughly 55 GB for weights in 16-bit precision, meaning deployment calls for substantial computing capacity rather than a typical gaming PC. Nvidia also says Flora can label up to 200,000 hours of video per day on 512 GB300 GPUs; that is a company-reported capacity figure, not evidence of independently verified production throughput.
Cost and Benchmark Caveats
The $0.72 figure is not an invoice for a completed Flora production run. It relies on an assumed GPU rate and modeled throughput, and excludes storage and overhead. Nvidia gives a range of $0.51 to $1.26 per video hour and says the production GPU-hour denominator has not been verified. The comparison also puts this modeled estimate beside Astra API usage recorded on a small test set; its ratio is about 122.5, rounded by Nvidia to approximately 125 times.
The performance picture is also incomplete. Nvidia’s reported overall score favors Flora on the selected episodes, but Astra leads on several robot-video and captioning measures. Most tests were run by Nvidia, and the provided material does not establish how results would transfer to different video collections, tasks, hardware or commercial operating conditions. The Flora reward model is not included in the 125-times cost comparison.
It is also unclear whether Nvidia will offer commercial terms, what a company would pay to use a comparable system legally in production, or what cost and quality would look like at sustained scale. Those questions cannot be answered by the research license or the benchmark estimate alone.
Commercial Terms and Real-World Tests
The next useful evidence would be independently replicated tests on a broader mix of robot footage, with the GPU hardware, throughput, utilization and full operating costs reported. Such tests could show whether Flora’s lower modeled captioning expense holds outside Nvidia’s selected benchmark and how its task-specific strengths and weaknesses affect real datasets.
For companies, the immediate next step is to review the license before any intended use; the stated terms do not authorize commercial production use. Whether Nvidia introduces separate commercial terms is not specified in the material. Until that is clear, the reported cost gap is best treated as a comparison for research and evaluation, not a confirmed saving available to businesses.
Key Questions
What is Nvidia Flora?
Flora is a set of three video-labeling models from Nvidia’s GEAR robotics lab. They identify action intervals, write captions and score robot performance.
Is Flora really 125 times cheaper than GPT-6 Astra?
Nvidia’s comparison implies a ratio of about 122.5 to one, rounded to approximately 125 times. It compares Astra API usage at $88.49 per video hour with Flora’s modeled $0.72 GPU cost, not two measured production bills.
Can a company use Flora commercially?
The stated NVIDIA OneWay Noncommercial License allows research or evaluation and restricts commercial use. The source material does not say whether Nvidia will offer separate commercial terms.
Does Flora outperform GPT-6 Astra on every test?
No. Nvidia reports Flora ahead on the overall score and on several measures, including human-video examples and partial successes. Astra scores better on robot videos, caption accuracy and rubric scoring.
What does the 125-times comparison include?
It compares the cost of captioning on 50 benchmark episodes totaling about 75 minutes of video. Flora’s reward model is not included, and the modeled cost excludes storage and overhead.
Source: ThorstenMeyerAI.com
