🔍 Read the full analysis: Nvidia’s Flora: What The 125× Robot Video Labeling Savings Leave Out on ThorstenMeyerAI.com
Get the latest gadgets 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 robot videos, and estimated captioning costs at $0.72 per video hour versus $88.49 for GPT-6 Astra. The roughly 122.5-to-1 ratio compares a modeled GPU budget with measured API usage; the benchmark is limited, results vary by task, and Flora’s license restricts use to research or evaluation.
Nvidia’s GEAR robotics lab has released Flora, a set of three models designed to label video used in robot training, and estimates its captioning pipeline costs $0.72 per video hour compared with $88.49 for GPT-6 Astra. The resulting roughly 122.5-to-1 ratio, rounded by Nvidia to about 125×, is not a like-for-like measurement: Flora’s cost is a modeled GPU estimate, while Astra’s is based on recorded API use.
Flora consists of three video models built on Alibaba’s Qwen3.5: a 9-billion-parameter model that identifies action intervals, a 27-billion-parameter captioner that describes those intervals, and a 27-billion-parameter rewarder that scores robot performance. The rewarder can rate an episode overall, label steps as normal, suboptimal or failed, or assess a video against a supplied rubric. Nvidia says the system was used as a data engine for its GR00T-2 robot models.
The cost comparison in Nvidia’s paper uses 50 episodes totaling 4,483 seconds, or about 75 minutes, from the company’s DenseCaptionBench. Astra’s $88.49 per video hour is based on reported API usage at public rates from October 4, 2026. Flora’s $0.72 figure is calculated from an assumed rate of $10 per GPU-hour for eight GB300 GPUs and modeled throughput. Nvidia’s paper says the production GPU-hour denominator has not been verified; its modeled scenarios range from $0.51 to $1.26 per hour, excluding storage and other overhead.
The ratio is therefore about 122.5 to 1 using the stated central estimates, not an independently measured production saving. The comparison covers captioning, not the reward model. The source report also notes that 106 of 7,062 Astra attempts had no reported usage, making the API total a known-usage subtotal. Nvidia’s paper reports aggregate benchmark scores of 57.1% for Flora and 53.7% for Astra on the tested episodes, but the report cautions that results differ across tasks and video types.
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 robot developers connect what a machine sees with what it should do, and identify where an action failed. Reducing the expense of generating those labels could make it easier to filter and organize large training datasets. Nvidia says Flora can process up to 200,000 video hours per day using 512 GB300 GPUs; that is a company-reported capacity figure, not evidence that customers are currently processing data at that scale.
The distinction between a lower estimated labeling cost and a lower total cost of building robots matters. Even if Flora’s modeled economics hold under suitable conditions, developers still need to gather useful video, run physical robot tests, provide computing capacity and address safety. The models also require substantial hardware: the source report estimates each 27B model’s 16-bit weights at roughly 55 GB, placing local inference beyond typical consumer computers.
For businesses, the noncommercial license is a direct constraint. The models are available for research or evaluation, but the source material says commercial use is not permitted under the stated license. That makes Flora useful as a research resource or comparison point, but not automatically a free production tool. The paper’s reported quality is mixed, too: an aggregate score does not establish that Flora is better for every robot, dataset or labeling task.
NVIDIA GPU workstation for AI video processing
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 recordings of robots and people performing tasks, such as placing objects or folding fabric. Training systems can use timestamped captions to identify what happened and when, along with process and outcome scores that help evaluate individual steps or an entire attempt. Flora separates these functions across its segmenter, captioner and rewarder.
According to the source report’s account of the paper and model cards, Flora was trained on fewer than 10,000 hours of human and robot video. Nvidia’s evaluation compares it with several other models, but most testing was run by Nvidia itself. The cited results include areas where Flora performs well, such as human-video captioning and partial-success judgments, and areas where Astra reportedly scores higher, including robot-video caption accuracy and rubric scoring. Those differences make the task-by-task results more informative than a single headline score.
The paper’s benchmark sample is 50 fixed episodes selected for Nvidia’s DenseCaptionBench, divided among human hands, two-finger grippers and dexterous robot hands. The source report also cites separate public benchmarks for segmentation and reward assessment. These evaluations provide evidence about the tested material, but they do not by themselves establish performance across the full range of real-world robot footage.
“The production GPU-hour denominator has not been verified.”
— Nvidia, in the Flora paper’s cost discussion
NVIDIA GB300 GPU server for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Cost Estimate Cannot Show
The headline comparison does not establish the cost of operating Flora in a commercial production environment. The GPU rate is an assumption, throughput is modeled, and the paper says its production GPU-hour denominator has not been verified. The scenarios also omit storage and overhead. It remains unclear how costs would change with different hardware availability, utilization, video quality, workloads or implementation choices.
The benchmark is small and was designed by Nvidia, while the source report says Nvidia conducted most of the tests. The available material does not establish whether the results generalize to other robot systems or substantially larger and more varied datasets. The aggregate quality score also does not resolve the reported differences among human video, robot video, caption accuracy and reward scoring.
Commercial use is another unresolved question for prospective adopters: the stated license is noncommercial, and the source material does not describe a separate commercial license or its terms. Nvidia’s ability to use the models commercially does not, on the information provided, grant the same rights to outside companies.
robotics development kit with camera and sensors
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Tests Needed Beyond the Paper
The next useful evidence would be independently reproduced cost and quality tests on larger, varied datasets, with measured GPU use and the full operating costs reported. Such tests could clarify whether the estimated savings persist outside Nvidia’s selected benchmark and under different deployment conditions.
Companies considering Flora will also need to check the license and hardware requirements before using the models in their workflows. Further details about commercial permissions, real-world throughput and performance on additional robot tasks would help determine whether Flora can move from research evaluation into broader deployment. No such results or licensing changes are confirmed in the material provided.
high capacity NVMe SSD for AI video datasets
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Nvidia Flora?
Flora is a group of three robot-video models: one segments videos into actions, one writes captions for those actions, and one scores robot performance.
Is the 125× saving a measured production result?
No. Nvidia’s roughly 122.5-to-1 calculation compares a modeled $0.72 GPU cost with $88.49 in reported GPT-6 Astra API usage. The paper says the production GPU-hour denominator has not been verified.
What videos were used in the comparison?
The reported cost and quality comparison used 50 Nvidia DenseCaptionBench episodes totaling 4,483 seconds, or about 75 minutes. The sample included human hands and two types of robot hands.
Can companies use Flora commercially?
The source report says all three models use the NVIDIA OneWay Noncommercial License, limited to research or evaluation purposes. It does not provide terms for a separate commercial license.
Does Flora outperform GPT-6 Astra on every task?
No. The source report describes mixed results: Flora leads on some human-video and partial-success tests, while Astra leads on some robot-video captioning and rubric-scoring measures.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
