🔍 Read the full analysis: Nvidia’s Flora Cuts Robot Video Labeling Costs—Here’s The License Fine Print 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 has released Flora, three models for segmenting, describing and scoring robot video. Nvidia estimates labeling costs at $0.72 per video hour, compared with $88.49 for GPT-6 Astra in its test, but the figures use different cost methods and Flora’s license permits research or evaluation only.
Nvidia’s GEAR robotics lab has released Flora, a three-model suite designed to mark actions in video, describe them and score robot performance. Nvidia says its labeling pipeline costs an estimated $0.72 per video hour, compared with $88.49 for GPT-6 Astra in a test—but the Flora figure is modeled, and its license allows research or evaluation, not commercial use.
Flora consists of Flora-TempSeg, a 9-billion-parameter model that identifies action intervals; a 27-billion-parameter Flora-Captioner, which writes instructions for those intervals; and a 27-billion-parameter Flora-Rewarder, which scores an episode or grades steps as normal, suboptimal or failure. Nvidia says the models were built on Alibaba’s Qwen3.5 models and used as a data engine for its GR00T-2 robotics model.
The cost comparison comes from 50 episodes in Nvidia’s DenseCaptionBench, amounting to 4,483 seconds of video, or about 75 minutes. Nvidia reports $88.49 per video hour for GPT-6 Astra based on API usage, while Flora’s $0.72 estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput. The paper gives Flora scenarios ranging from $0.51 to $1.26 per hour, excluding storage and overhead, and says the production GPU-hour denominator has not been verified. The comparison covers captioning, not Flora’s reward model.
On the same episodes, the paper 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 judgments, while Astra performed better on robot videos, caption accuracy and rubric scoring. Nvidia conducted most of the reported evaluations, so the results are not an independent assessment of performance across all robot datasets.
The three models are available through Hugging Face, with inference code in their repositories. Their model cards specify the NVIDIA OneWay Noncommercial License, which limits use to research or evaluation and applies the same restriction to derivatives. That means the release does not grant companies permission to run Flora as a commercial labeling service or build a commercial product from it under the stated terms.
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 Case for Robot Data
Video labeling is one step in preparing the large datasets used to train and evaluate robot systems. If Flora’s cost estimate holds at scale, it could lower the expense of adding time-aligned instructions and performance scores to video, making it easier for researchers to process more examples or examine failures in greater detail. Nvidia says Flora can label data for its GR00T models at a rate of up to 200,000 hours of video per day using 512 GB300 GPUs; that is a company-reported capacity, not an independently verified production result.
The headline price difference is not a like-for-like measurement: it compares a modeled GPU budget with recorded API charges on a small, Nvidia-selected test set. The results also show a trade-off in quality rather than a clear win across tasks. For businesses, the license is a separate constraint: the models may help as a research reference, but commercial deployment would require different permission or a separately licensed alternative. Lower labeling costs alone do not resolve the expense of collecting robot data, building hardware or demonstrating safe operation.
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
Samsung 27-inch Odyssey OLED G6 QD-OLED QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Three Models, Three Labeling Jobs
Robot-learning systems use video of both robots and people performing tasks. According to Nvidia’s paper, timestamped captions identify what happens and when; process rewards grade steps along the way; and outcome rewards score an entire attempt. These labels can help developers organize training examples and identify where a robot’s behavior falls short.
Flora divides those tasks across its models: TempSeg finds the intervals, Captioner describes them, and Rewarder assigns scores. The model cards say the suite was trained on fewer than 10,000 hours of robot and human video. The paper’s cost comparison is specifically about captioning, so it should not be read as a demonstrated price for every part of a robot-data pipeline.
Limits of the Cost Comparison
The reported 122.5-to-1 cost ratio, rounded by Nvidia to approximately 125 times, depends on its GPU-price assumption and modeled throughput. Nvidia says the production GPU-hour denominator is unverified; the estimate also excludes storage and overhead. The Astra figure reflects reported API usage, with 106 of 7,062 attempts lacking usage data, according to the source paper. The test covers about 75 minutes of video from 50 episodes, not a broad production workload.
It is also unclear how the models will perform on video collected by other organizations, with different cameras, tasks or robot designs. The benchmark results are mixed, and most tests were run by Nvidia. The supplied source does not establish whether Nvidia will offer commercial licensing, or whether the license terms will change. Flora’s stated noncommercial terms do not authorize commercial use unless Nvidia grants separate permission.
Independent Tests and License Terms
The next useful evidence would be independent evaluations on larger and more varied robot-video collections, alongside measured costs from actual deployments. Those results could show whether Flora’s modeled savings persist when infrastructure, storage and operational overhead are included, and whether its task-specific quality differences matter in practice.
Companies considering the models will need to review the license and seek clarification or separate authorization before any commercial use. Until then, Nvidia’s release gives researchers a set of tools to test and compare, while the cost advantage remains an estimate rather than a verified production bill.
Key Questions
What does Nvidia Flora do?
Flora includes models that segment video into action intervals, caption those intervals and score robot actions or full episodes.
Is the $0.72-per-hour cost a measured production price?
No. Nvidia describes it as a modeled estimate based on eight GB300 GPUs at an assumed $10 per GPU-hour. The paper says the production GPU-hour denominator has not been verified and excludes storage and overhead.
Does Flora beat GPT-6 Astra on every benchmark?
No. Nvidia reports stronger Flora results on some areas, including human videos and partial-success judgments, while Astra leads on robot videos, caption accuracy and rubric scoring.
Can a company use Flora commercially under its listed license?
The stated NVIDIA OneWay Noncommercial License permits research or evaluation only and extends that restriction to derivatives. Commercial use would require separate permission or different licensing.
Source: ThorstenMeyerAI.com
