🔍 Read the full analysis: A License Check For Nvidia Flora’s 125× Robot Video Labeling Cost Claim 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. Nvidia estimates captioning costs $0.72 per video hour, compared with $88.49 for GPT-6 Astra, but the comparison mixes a modeled GPU cost with measured API usage; Flora’s license also restricts use to research or evaluation.
Nvidia’s GEAR robotics lab has released Flora, a set of three models for labeling robot video, and estimates that its captioning system costs $0.72 per video hour versus $88.49 for GPT-6 Astra. The claimed roughly 125-fold cost gap is not a like-for-like bill comparison: Nvidia modeled Flora’s GPU cost while measuring Astra’s API usage, and Flora’s license limits use to research or evaluation.
Flora includes a 9-billion-parameter temporal segmenter that identifies action intervals, a 27-billion-parameter captioner that writes instructions for those intervals, and a 27-billion-parameter reward model that scores robot attempts. Nvidia says the models were used to label data for its GR00T robot models. The project paper and model cards say Flora was trained on fewer than 10,000 hours of human and robot video.
The paper’s cost test used 50 episodes totaling 4,483 seconds, or about 75 minutes, from Nvidia’s DenseCaptionBench. Nvidia reports Astra’s measured API usage at $88.49 per video hour. Flora’s $0.72 figure is a modeled estimate based on eight GB300 GPUs priced at an assumed $10 per GPU-hour; the paper gives a range of $0.51 to $1.26 and excludes storage and overhead. Dividing the two headline figures gives about 122.5, rounded by Nvidia to approximately 125 times.
On the same test set, Nvidia reports Flora scoring 57.1% and Astra 53.7%. The paper’s broader results are mixed: Flora performs better on human-video examples, action boundaries and some partial-failure assessments, while Astra leads on robot-video caption accuracy and rubric scoring. Nvidia conducted most of the evaluations itself, so independent replication would help establish how well the results generalize.
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 Claim Has Two Caveats
Video annotation is a significant expense in building robot training datasets: systems need timestamps, descriptions of actions and assessments of whether steps succeeded. If Flora can provide useful labels at a much lower cost, it could reduce one part of the data-processing burden for robotics developers. Nvidia says it can process as much as 200,000 hours of video per day using 512 GB300 GPUs, though that is a vendor-reported capacity figure, not an independently verified production result.
The headline savings need careful interpretation. Astra’s figure is based on reported API usage, while Flora’s is a modeled GPU budget using an assumed rental price and throughput. Nvidia’s paper says the production GPU-hour denominator has not been verified. The estimate also covers captioning, not all three Flora models or the full cost of operating a labeling pipeline. It does not establish that a company would see a 125-fold reduction in total costs.
Access is another limit for businesses. The models are distributed under the NVIDIA OneWay Noncommercial License, which the source describes as allowing research or evaluation only, with derivatives carrying the same restriction. That makes Flora a potential research tool and technical reference, not a license-cleared, free commercial labeling service. Companies would need to review the license and seek separate authorization before relying on it in production.
Products Worth Considering
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
As an affiliate, we earn on qualifying purchases.
Three Models, Three Labeling Tasks
Robot-learning datasets contain recordings from cameras mounted on robots, as well as videos of people performing tasks. For training, a recording may need to be divided into action segments, paired with a concise instruction and graded for success or failure. Flora separates these functions: TempSeg finds action boundaries, Captioner describes them, and Rewarder scores an episode or individual steps.
The models are built on Alibaba’s Qwen3.5 family, according to the project materials. Nvidia’s reported cost comparison concerns the captioning benchmark; the reward model is not included in the 125-times figure. The benchmark used a fixed collection of Nvidia-selected episodes, split among human hands, two-finger grippers and dexterous robot hands. That design provides a defined basis for comparison, but results on it do not by themselves establish performance across other datasets, tasks or deployment settings.
The model cards identify H100 hardware for testing. The source estimates that each 27-billion-parameter model requires roughly 55 GB for weights at 16-bit precision, underscoring that local use requires data-center-class resources rather than a typical consumer computer. Nvidia says the models are available through Hugging Face repositories, which include inference code.
“approximately 125×”
— Nvidia, describing the estimated comparison
Costs and Performance Need Replication
The available source material does not establish Flora’s actual production cost per video hour. Its estimate depends on assumed GPU pricing and throughput, and excludes storage and other operating overhead. Astra’s number is reported API usage from a small, fixed test set; the article notes that 106 of 7,062 attempts had no usage reported, making the figure a known-usage subtotal.
It is also unclear how the cost and quality results would change on video collections outside Nvidia’s benchmark, or when a company includes data preparation, review, infrastructure and maintenance. The reported scores show strengths and weaknesses across tasks rather than a uniform quality advantage. The source provides no independent evaluation of the benchmark results or confirmation of the stated daily processing capacity.
The license’s practical application to particular business arrangements may require legal review. The source says commercial use is not allowed under the listed license and that derivatives inherit the restriction, but does not describe any separate commercial licensing terms from Nvidia.
Independent Tests and License Terms
The next useful evidence would be independent testing on a broader set of robot and human videos, along with measured end-to-end operating costs that include GPU time, storage and review. Such results could show whether Flora’s benchmark performance and estimated savings hold beyond Nvidia’s selected test episodes.
Organizations considering the models should check the current model cards and license before use, especially if their work could become part of a commercial product or service. Nvidia’s paper and repositories provide the stated technical details; the source material does not identify a date for a broader independent evaluation, a verified production-cost audit or an announced commercial license.
Key Questions
What is Nvidia Flora?
Flora is a three-model system for labeling robot video: it segments actions, captions them and scores robot performance.
Is Flora really 125 times cheaper than GPT-6 Astra?
Nvidia reports a ratio of about 122.5, rounded to approximately 125, using $0.72 per video hour for Flora and $88.49 for Astra. Flora’s cost is modeled from assumed GPU pricing and throughput; Astra’s is based on measured API usage, so the figures are not directly equivalent.
Does the 125-times comparison include all three Flora models?
No. The source says the comparison concerns captioning only. It does not include the Flora reward model or the full cost of running a labeling operation.
Can companies use Flora commercially?
The source describes the models as released under the NVIDIA OneWay Noncommercial License, for research or evaluation. It says derivatives inherit that restriction; companies should review the license and ask Nvidia about any separate commercial permissions.
How strong are Flora’s benchmark results?
Nvidia reports a 57.1% score for Flora versus 53.7% for Astra on its 50-episode test. The paper reports advantages for Flora on some human-video and partial-failure tasks, while Astra performs better on some robot-video captioning and rubric-scoring measures. The results have not been independently verified in the supplied material.
Source: ThorstenMeyerAI.com
