🔍 Read the full analysis: Nvidia Flora Makes Robot Video Labeling Cheaper. What’s The License Catch? on ThorstenMeyerAI.com
Get hardware and tech essentials 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 versus $88.49 for GPT-6 Astra on its test, but Flora’s figure is modeled, its quality results are mixed, and its license limits use to research or evaluation.
Nvidia’s GEAR robotics lab has released Flora, a set of three models designed to label robot-training videos, and estimates that its captioning pipeline costs $0.72 per video hour compared with $88.49 for GPT-6 Astra on the lab’s test. The headline saving is based on a modeled GPU cost, not a measured production bill, and Flora’s license permits research or evaluation only, not commercial deployment.
Flora includes a 9-billion-parameter segmenter that identifies when actions begin and end, a 27-billion-parameter captioner that describes each segment, and a 27-billion-parameter rewarder that scores a robot’s progress or outcome. The models are built on Alibaba’s Qwen3.5 family. Nvidia says it used Flora as a data engine for its GR00T-2 robotics model and trained the models on fewer than 10,000 hours of human and robot video.
The cost comparison in the Flora paper uses 50 episodes totaling 4,483 seconds, or about 75 minutes of video. GPT-6 Astra’s $88.49-per-hour figure comes from reported API usage at public rates; Flora’s $0.72 estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput. Nvidia’s paper gives Flora scenarios ranging from $0.51 to $1.26 per video hour, excluding storage and other overhead. It also says the production GPU-hour denominator has not been verified. The comparison concerns captioning, not the cost of every Flora model or a full robot-training pipeline.
On the same test episodes, the paper reports a Flora score of 57.1%, compared with 53.7% for GPT-6 Astra, 38.3% for Gemini 3.8 Flash and 30.4% for Claude Opus 5. The paper reports Flora performing better on human videos, action boundaries, failed steps and partial success, 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.
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 Meets Its License
Video labeling is a labor- and compute-intensive step in building robot datasets. Models that can identify actions, describe them and grade outcomes may make it easier to sort large collections of recordings and find examples of success or failure. If the cost estimates hold at scale, lower labeling costs could reduce one expense in developing robot-learning systems.
But the commercial-use restriction changes what the release offers companies. Under the NVIDIA OneWay Noncommercial License, the models are for research or evaluation, and derivatives inherit that restriction. A business cannot treat Flora as a free production labeling service under the stated terms. The release may still help researchers test approaches and give companies a technical reference, but commercial use would require a different permission or a separately licensed alternative.
The result also does not remove other barriers to robotics. The estimate covers a specific labeling task and omits overhead; it says nothing by itself about the cost or availability of robot video, hardware, deployment or safety work. The benchmark’s mixed results mean organizations would also need to check whether Flora performs well on their particular cameras, tasks and data.
Products Worth Considering
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
Apple iPad Air 13-inch (M4), 128GB, Wi-Fi 7, Space Gray
Amazon Kindle Paperwhite Signature Edition (12th Generation, 2024), 32 GB, Metallic Black
As an affiliate, we earn on qualifying purchases.
How Flora Labels Robot Videos
Robot-learning systems use video to learn how actions unfold. The Flora paper describes three kinds of labels: timestamped captions that connect instructions to parts of a recording; process rewards that classify short steps as normal, suboptimal or failures; and outcome rewards that score an entire attempt. These labels can help researchers organize training data and identify where an action went wrong.
Flora’s segmenter finds action intervals, its captioner writes an instruction for each interval using surrounding video for context, and its rewarder can score an episode on a 1-to-5 scale, grade steps at half-second intervals or apply a user-provided rubric. The model cards say the captioner can use one camera or a main camera with wrist cameras. Nvidia says its cost comparison was against GPT-6 Astra, Gemini 3.8 Flash and Claude Opus 5, but the headline 125-times comparison is specifically the ratio of Astra’s reported $88.49 cost to Flora’s modeled $0.72 estimate.
“The cost figure is a modeled estimate, the quality picture is mixed, and ‘open’ here means free for research, not for business.”
— Thorsten Meyer, ThorstenMeyerAI.com
Open Questions on Cost and Performance
The $0.72 figure is an estimate based on assumed GPU pricing and throughput, not a measured bill from a production deployment. Nvidia’s paper says the GPU-hour denominator has not been verified, and the estimate excludes storage and overhead. The test uses 50 episodes selected for Nvidia’s DenseCaptionBench; results on other datasets or larger, varied production workloads have not been established by the reported comparison.
The quality evidence is also limited. Flora’s overall score was higher than Astra’s on this test, but Astra led on robot videos, caption accuracy and rubric scoring. Nvidia ran most of the evaluations, and the supplied material does not establish how the models perform across independent tests or real-world deployments. The license material provided states the noncommercial restriction, but it does not establish whether Nvidia will offer separate commercial terms or services.
Independent Tests and Commercial Terms
The next useful evidence will be independent evaluations on robot-video datasets beyond Nvidia’s benchmark, with comparable reporting of quality, compute, throughput and total operating costs. Practical tests will also need to show how Flora handles different robots, camera setups and tasks, and whether its results remain consistent at larger volumes.
For organizations considering use, the immediate step is to review the applicable license before using the models or derivatives in a business workflow. Nvidia has not, in the material cited here, announced commercial licensing terms or a production service. Until those details or independent deployment measurements emerge, the cost comparison is best read as a promising modeled estimate for research—not a confirmed price for commercial robot-data labeling.
Key Questions
What is Nvidia Flora?
Flora is a set of three video models from Nvidia’s GEAR robotics lab: one segments actions in video, one captions those segments, and one scores robot actions or outcomes.
Is Flora really 125 times cheaper than GPT-6 Astra?
That ratio comes from dividing Astra’s reported $88.49 per video hour by Flora’s modeled estimate of $0.72. Astra’s figure is based on API usage, while Flora’s uses assumed GPU costs and throughput; it is not a measured, like-for-like production bill.
Can a company use Flora commercially?
The stated NVIDIA OneWay Noncommercial License allows use for research or evaluation only, and derivatives inherit that restriction. The source material does not identify separate commercial terms.
Did Flora outperform GPT-6 Astra on every test?
No. Nvidia’s paper reports a higher overall score for Flora on its test episodes, but Astra scored better on robot videos, caption accuracy and rubric scoring. The evaluation was largely conducted by Nvidia.
Source: ThorstenMeyerAI.com
