🔍 Read the full analysis: Can The License Keep Up With Nvidia Flora’s Robot Video Labeling Cost Savings? on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Nvidia’s GEAR lab has released Flora, a three-model system for labeling robot video, and estimates its captioning cost at $0.72 per video hour versus $88.49 for GPT-6 Astra on a selected test set. The comparison is partly modeled, quality results are mixed, and Flora’s NVIDIA OneWay Noncommercial License restricts use to research or evaluation—not commercial production.
Flora comprises three video models: a 9-billion-parameter segmenter that identifies when actions begin and end, a 27-billion-parameter captioner that describes each action, and a 27-billion-parameter reward model that scores performance. Nvidia says the system was used as a data engine for its GR00T-2 robot models. The models were trained on fewer than 10,000 hours of human and robot video, according to their model cards.
The cost comparison in Nvidia’s paper uses 50 episodes totaling 4,483 seconds, or about 75 minutes. GPT-6 Astra’s $88.49 per video hour is based on reported API usage at public rates. Flora’s $0.72 figure is a projection based on eight GB300 GPUs priced at an assumed $10 per GPU-hour and modeled throughput. Nvidia gives a Flora range of $0.51 to $1.26 per hour, excluding storage and other overhead; it says the production GPU-hour denominator has not been verified. The paper’s approximately 125-times ratio is calculated from $88.49 divided by $0.72.
Quality results do not show Flora winning every task. On Nvidia’s selected DenseCaptionBench episodes, Flora scored 57.1%, compared with 53.7% for Astra. The paper reports Flora performing better on human videos, action boundaries, and partial-success judgments, while Astra did better on robot videos, caption accuracy, and rubric scoring. Nvidia conducted most of the reported tests itself. The 125-times cost comparison applies to captioning; it does not include Flora’s reward model.
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, Limited Use
Video annotation is a recurring expense in training robotics systems: developers need timestamps, descriptions of actions, and judgments about whether steps succeeded. If Flora’s modeled cost translates to production, the difference could make it less expensive to process large collections of footage and reduce one barrier to building robot-training datasets.
But the license changes the practical value for companies. All three models use the NVIDIA OneWay Noncommercial License, which limits use to research or evaluation and applies the restriction to derivatives. A company cannot treat the published models as a ready-made, free commercial labeling pipeline under those terms. Nvidia’s own right to use Flora commercially does not grant the same permission to other users.
The headline savings also need to be read alongside the benchmark limits. The test was small, chosen by Nvidia, and compares a modeled cost for running Flora with an API bill for Astra. It does not establish the cost of a commercial-scale service, or show that Flora is more accurate across the full range of robot footage. For users and developers, cheaper labeling is relevant, but it does not remove other constraints such as collecting useful data, building safe robots, or paying for hardware.
NVIDIA GPU workstation for AI video processing
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Three Models for Robot Video
Robot-learning systems use video of people and machines carrying out tasks, such as placing an object or folding fabric. To make that footage useful for training and evaluation, developers may need timestamped action captions, step-by-step reward labels, and scores for the overall attempt. These labels help organize data and identify successful or failed behavior.
Flora’s models are built on Alibaba’s Qwen3.5 family, according to Nvidia’s paper and model cards. The segmenter finds action intervals; the captioner describes them in English using surrounding video for context; and the rewarder can score an episode from 1 to 5, classify half-second segments as normal, suboptimal, or failure, or apply a user-provided rubric. The inference code is included in the models’ Hugging Face repositories.
Nvidia’s published testing used an H100 system on Linux. The source material estimates that each 27-billion-parameter model requires roughly 55 GB for weights at 16-bit precision, making the models suited to data-center hardware rather than an ordinary gaming PC. That infrastructure requirement is separate from the paper’s estimated per-hour processing cost.
““The production GPU-hour denominator has not been verified.””
— Nvidia, in the Flora paper
NVIDIA data center GPU server for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Cost and Benchmark Limits
The $0.72-per-hour figure is not a measured production bill. It depends on an assumed GPU rental price and modeled throughput, and excludes storage and overhead. Nvidia reports a range rather than a single guaranteed cost and acknowledges that the production GPU-hour denominator has not been verified. The comparison also uses 106 of 7,062 Astra attempts with reported usage, so the paper describes the API total as a known-usage subtotal.
It remains unclear how the models would perform and cost across broader, independently selected datasets and real production workloads. The reported quality score covers a small set of episodes selected by Nvidia, and results differ by task and video type. The supplied source material does not establish whether the license may change, whether separate commercial terms are available, or how Flora’s costs compare once a company accounts for deployment, engineering, and data handling.
robotics video annotation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Commercial Terms and Real-World Tests
The next useful evidence would be independent testing on larger and more varied robot-video datasets, alongside published production measurements of throughput, GPU use, and total operating cost. Those results would help establish whether the modeled savings hold beyond Nvidia’s selected test and assumptions.
Companies considering Flora will also need to review the license before use and determine whether research or evaluation fits their intended work. Until Nvidia publishes verified production economics or different licensing terms, Flora is best understood as a research and evaluation release with a potentially low-cost labeling approach—not as a commercially cleared alternative to paid annotation services.
AI model training workstation with multiple GPUs
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Nvidia Flora?
Flora is a set of three models from Nvidia’s GEAR robotics lab: a video action segmenter, a captioning model, and a model that scores robot performance.
Is Flora really 125 times cheaper than GPT-6 Astra?
Nvidia’s paper reports an estimated Flora cost of $0.72 per video hour and $88.49 for GPT-6 Astra on its 50-episode test, a ratio of about 122.5 to one. Flora’s figure is modeled using assumed GPU costs and throughput, while Astra’s is based on reported API usage; the comparison is not a verified production-cost measurement.
Can businesses use Flora commercially?
The models are released under the NVIDIA OneWay Noncommercial License, which says they are for research or evaluation purposes only. The source material says derivatives inherit the restriction, so the release does not grant businesses general commercial-use rights.
Does Flora outperform GPT-6 Astra on every test?
No. Nvidia reports Flora ahead on some measures, including human-video performance and partial-success judgments, while Astra leads on robot videos, caption accuracy, and rubric scoring. The results come from Nvidia’s tests and do not establish performance on all robot-video datasets.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
