🔍 Read the full analysis: How To Weigh Nvidia Flora’s Robot Video Labeling Savings Against Its License 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, and estimates captioning costs of $0.72 per video hour versus $88.49 for GPT-6 Astra. That comparison pairs a modeled GPU cost with measured API usage, and Flora’s NVIDIA OneWay license permits research and evaluation, not commercial deployment.
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 to run, compared with $88.49 for GPT-6 Astra on a test set. The comparison is not like-for-like: Nvidia’s figure is a modeled GPU budget, while Astra’s is based on measured API usage, and Flora’s license restricts use to research or evaluation, excluding commercial deployment.
Flora comprises three models built on Alibaba’s Qwen3.5 family. Flora-TempSeg, with 9 billion parameters, identifies the time intervals in which actions occur. The 27-billion-parameter Flora-Captioner writes an instruction for each interval, while Flora-Rewarder scores an entire episode or grades short steps as normal, suboptimal or failure. Nvidia says the models were used as a data engine for its GR00T-2 robot models and were trained on fewer than 10,000 hours of human and robot video.
The cost comparison uses 50 episodes totaling 4,483 seconds, or about 75 minutes, from Nvidia’s DenseCaptionBench. The paper reports GPT-6 Astra API usage costing $88.49 per video hour. Flora’s estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput, producing a figure of $0.72 per hour, with modeled scenarios from $0.51 to $1.26. Nvidia divides $88.49 by $0.72 to reach a ratio of about 122.5 to one, rounded in the paper to approximately 125 times.
That estimate applies to captioning only; it does not include the Flora-Rewarder’s scoring work. The paper also records 106 of 7,062 Astra attempts without reported usage, so its API-cost total is a known-usage subtotal. Nvidia states that the production GPU-hour denominator behind its own estimate has not been verified, and storage and overhead are excluded.
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 Depends on Deployment
Video labeling is a labor- and compute-intensive part of building robot-learning datasets. Timestamped descriptions help teams sort and balance training data, while step-level and episode-level scores can support reinforcement learning and evaluation. If the modeled savings translate into production, lower labeling costs could let robotics teams process more video or reduce spending on this part of their pipelines.
For companies, however, the license changes the practical value of the release. Flora is available under Nvidia’s OneWay Noncommercial License, which allows research or evaluation and passes the same restriction to derivatives. The comparison can inform internal research and benchmarking, but it does not establish a free commercial labeling service. Businesses would need a separate permission or a different tool before using Flora in a commercial workflow.
The savings claim also should not be read as a full measure of robot-development costs. It addresses one labeling task, and even a cheaper labeling pipeline does not resolve the expense of collecting robot data, operating hardware or validating systems for safe use. The results are relevant to the economics of dataset preparation, not a guarantee that robots will become cheaper to build or deploy.
Products Worth Considering
Apple MacBook Pro 14-inch with M5 Max Chip (18-Core CPU, 20-Core GPU), 48GB RAM, 2TB SSD, Space Black
Apple MacBook Pro 14 (2021) with M1 Max, 64 GB RAM, 2 TB SSD, Silver
Apple MacBook Pro 2021 16‑inch with M1 Max, 32 GB RAM, 2 TB SSD, Space Gray (QWERTY English)
As an affiliate, we earn on qualifying purchases.
Three Models, Three Labeling Tasks
Robot-learning systems need more than raw recordings. A video of a robot moving an object can be annotated with when each action happens, a caption describing the action, and a score indicating whether the attempt succeeded. These labels help researchers locate useful examples and identify where a task went wrong. Flora divides those functions among its segmenter, captioner and rewarder rather than relying on one model for every output.
The released model cards describe the captioner as accepting one camera view or a main view with wrist cameras. Nvidia’s paper compares caption quality on a fixed benchmark, but its reported headline cost ratio concerns captioning, not all three Flora models. The paper’s testing used an Nvidia H100 for model inference; the cost estimate, by contrast, assumes a fleet of GB300 GPUs. Those distinctions matter when evaluating whether another organization could reproduce the reported price.
On the same 50-episode test set, Nvidia reports Flora at 57.1% quality and GPT-6 Astra at 53.7%. The broader benchmark results are mixed: Flora performs better on human-video examples, action boundaries and some partial-success cases, while Astra scores better on robot videos, caption accuracy and rubric scoring. Most tests were run by Nvidia, so the results are company-reported benchmark findings rather than an independent production audit.
“research or evaluation purposes only”
— NVIDIA OneWay Noncommercial License
Cost and Benchmark Limits
The central unresolved issue is whether Flora can achieve the modeled throughput and GPU cost in a real production environment. Nvidia’s paper says its GPU-hour denominator is unverified, and the estimate excludes storage and other overhead. The test set is relatively small, contains episodes selected for Nvidia’s benchmark, and does not show that other operators would see the same costs or quality.
The comparison also uses different cost methods: a measured API charge for Astra and an assumed rental rate and modeled throughput for Flora. The paper’s quality score does not settle which model is best for every task, since the reported results favor different systems on different video types and scoring measures. It remains unclear whether independent tests will reproduce the headline savings, how the models perform across larger and more varied datasets, or whether Nvidia will offer commercial licensing.
Independent Tests and Licensing
The next useful evidence would be independent evaluations using larger video sets and documented, end-to-end operating costs. Those tests would need to report GPU hardware, throughput, rental or ownership costs, storage, and the same labeling workload for each model. That would help distinguish a benchmark estimate from a reproducible production cost.
Organizations considering Flora should also review the license before building it into a workflow. Unless Nvidia changes the terms or grants permission, the published license limits use to research and evaluation. The source material does not identify a schedule for further benchmark releases, a commercial license, or a new cost audit, so no such next milestone has been announced.
Key Questions
What is Nvidia Flora?
Flora is a set of three models from Nvidia’s GEAR robotics lab: one segments action intervals in video, one captions those intervals, and one scores robot actions or full episodes.
Is the $0.72 cost a measured production price?
No. It is a modeled estimate based on eight GB300 GPUs, an assumed rental cost of $10 per GPU-hour and modeled throughput. The paper says the production GPU-hour denominator has not been verified; storage and overhead are excluded.
What does the approximately 125-times comparison cover?
It compares Nvidia’s modeled $0.72 per video hour for Flora captioning with GPT-6 Astra’s reported $88.49 per hour in API usage on a 50-episode test set. It does not include Flora’s reward model, and the two costs are calculated using different methods.
Can a company use Flora commercially?
The published NVIDIA OneWay Noncommercial License permits use for research or evaluation purposes only. Commercial use is not authorized by those terms, and derivatives inherit the restriction unless separate permission is obtained.
Does Flora outperform GPT-6 Astra on every test?
No. Nvidia reports higher overall quality for Flora on its 50-episode test and advantages on human videos, action boundaries and some partial-success cases. Astra performs better on robot videos, caption accuracy and rubric scoring in the reported comparisons.
Source: ThorstenMeyerAI.com
