🔍 Read the full analysis: The License Caveat Behind Flora’s 125× Cheaper Robot Video Labeling 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, describing and scoring robot videos, and estimates that its captioning costs about $0.72 per video hour. The roughly 125× comparison with GPT-6 Astra contrasts a modeled GPU cost with measured API use; Flora’s license also limits use to research or evaluation, not commercial production.
Nvidia’s GEAR robotics lab has released Flora, a set of three models for labeling robot videos, alongside an estimate that its captioning system costs about $0.72 per hour of video—roughly 125 times less than GPT-6 Astra in Nvidia’s comparison. The headline figure is not a measured production bill, however, and the models’ OneWay Noncommercial License restricts them to research or evaluation rather than commercial deployment.
Flora comprises a 9-billion-parameter temporal segmenter, which identifies when actions begin and end; a 27-billion-parameter captioner, which describes each segment; and a 27-billion-parameter rewarder, which scores overall attempts or individual steps. Nvidia says the models were built on Alibaba’s Qwen3.5 models and were used as a data engine for its GR00T-2 robot models. The model cards say training used fewer than 10,000 hours of human and robot video.
The cost comparison uses 50 episodes totaling about 75 minutes from Nvidia’s DenseCaptionBench. GPT-6 Astra’s $88.49 per video hour is based on reported API usage at public rates from October 4, 2026. Flora’s $0.72 is a modeled estimate based on eight GB300 GPUs priced at an assumed $10 per GPU-hour and projected throughput; Nvidia gives scenarios from $0.51 to $1.26 and excludes storage and overhead. The paper says the production GPU-hour denominator has not been verified. The comparison covers captioning, not Flora’s reward model.
On the same test episodes, Nvidia reports Flora at 57.1% quality and Astra at 53.7%. The broader results are mixed: Flora scores better on human videos and partial-success cases, while Astra leads on robot videos, caption accuracy and rubric scoring. Nvidia conducted most of the evaluations, so the results are not an independent assessment of performance across production datasets.
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 identify actions, locate failures and judge whether a task succeeded. Automating that work could reduce one cost in preparing data for robot training, particularly when teams need to process large video collections. Nvidia says Flora can label up to 200,000 hours of video a day using 512 GB300 GPUs; that is a company-reported capacity figure, not an independently verified production result.
The commercial license sharply narrows what businesses can do with the release. The models are available for research or evaluation only, and the source material says derivatives inherit that restriction. Nvidia itself may use the models commercially, but outside companies cannot treat the released weights as a free production labeling service under this license. For them, Flora is currently a research tool and a benchmark for comparing costs, rather than an unrestricted substitute for paid services.
Cost matters, but it is not the only constraint on robotics progress. Even substantially cheaper labeling would not itself supply the robot footage, reliable hardware or safety testing needed to build and deploy capable systems. The mixed benchmark results also mean that developers should not assume Flora performs better on every kind of video or scoring task.
Products Worth Considering
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
LG UltraGear 32GX850A-B 32-Inch 4K UHD Glossy OLED Gaming Monitor
Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Eucalyptus Green
As an affiliate, we earn on qualifying purchases.
Three Models for Robot Video
Robot-learning datasets can include recordings from robot cameras and human hands performing tasks. The Flora paper describes three labeling needs: timestamped captions that connect instructions to video intervals, step-level rewards that mark normal, suboptimal or failed actions, and scores for an entire attempt. These labels can help organize training data and evaluate behavior.
Flora separates those jobs across its models. The segmenter finds action boundaries, the captioner generates an English instruction for each interval, and the rewarder can score a full episode from 1 to 5, grade steps at half-second intervals, or apply a user-provided rubric. The captioner supports one camera or a main camera with wrist cameras. Nvidia’s model cards specify H100 test hardware; the article’s source estimates that each 27-billion-parameter model needs about 55 GB for weights in 16-bit precision, making the models suited to data-center systems rather than typical consumer computers.
The 125× figure is a ratio of $88.49 divided by $0.72, or about 122.5, rounded by Nvidia. The two sides have different evidence bases: Astra’s cost comes from observed API usage, while Flora’s comes from modeled GPU time and an assumed rental price. Nvidia also reports that 106 of 7,062 attempts had no usage reported, making the Astra figure a known-usage subtotal. The paper’s benchmark results and cost estimate should therefore be read as a limited comparison, not a guaranteed price or performance result for other workloads.
“research or evaluation purposes only”
— NVIDIA OneWay Noncommercial License
Cost and Performance Still Need Testing
The estimate does not establish what Flora would cost on a company’s own infrastructure or at sustained production scale. The assumed $10 per GPU-hour, projected throughput and unverified production GPU-hour denominator leave room for actual costs to differ; storage and overhead are excluded. The source also notes that Astra’s API total omits attempts where usage was not reported.
How well the models generalize beyond Nvidia’s selected episodes is also not established. The 50-episode comparison is small, and benchmark performance varies by video type and task. Independent testing across larger, varied datasets would help clarify whether Flora’s reported strengths hold outside Nvidia’s evaluations. The source provides no evidence that the 125× ratio applies to the rewarder or to complete labeling workflows.
The license’s application to particular business arrangements, and whether Nvidia might later change the terms or offer a separate commercial path, are not specified in the supplied material. The article’s source says derivatives inherit the noncommercial restriction, but does not detail licensing advice for specific uses.
Independent Tests and License Terms
The next useful evidence would be independent evaluations of Flora on larger robot-video datasets, alongside cost measurements from actual deployments that include hardware utilization, storage and operational overhead. Those tests could show whether the modeled savings persist and where the quality trade-offs matter most.
Companies considering the models should review the OneWay Noncommercial License before using the weights or derivatives, and should not assume research access grants commercial rights. Nvidia has not, in the source material, announced a commercial license or a revised cost estimate. Until it does, the reported price advantage remains a modeled comparison and the release’s stated use remains noncommercial.
Key Questions
What is Flora?
Flora is a set of three Nvidia GEAR models for finding action intervals in video, writing captions for those intervals and scoring robot attempts or steps.
Is Flora really 125 times cheaper than GPT-6 Astra?
Nvidia’s rounded comparison is based on $0.72 per video hour for Flora and $88.49 for Astra. Flora’s figure is modeled using assumed GPU pricing and projected throughput; Astra’s is based on reported API usage on 50 test episodes. The ratio is not a verified production-cost comparison.
Can companies use Flora commercially?
The supplied source says the models use the NVIDIA OneWay Noncommercial License, for research or evaluation only, with derivatives inheriting that limit. Companies should not treat the release as permission for commercial production use.
Does Flora outperform Astra on every test?
No. Nvidia reports Flora ahead on the selected overall quality score and on some human-video and partial-success tests, while Astra leads on robot videos, caption accuracy and rubric scoring. Most evaluations were run by Nvidia.
Source: ThorstenMeyerAI.com
