🔍 Read the full analysis: A Closer Look At Nvidia Flora’s 125× Robot Video Labeling Cost Reduction on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Nvidia’s GEAR lab released Flora, three models for segmenting, captioning and scoring robot videos. Nvidia estimates captioning could cost $0.72 per video hour, about 122.5 times below a measured GPT-6 Astra API cost, but the comparison uses a modeled GPU budget, a 50-episode test set and a noncommercial license. Benchmark results are mixed, and the reward models are not included in the cost ratio.
Nvidia’s GEAR robotics lab has released Flora, a set of three models designed to segment, describe and score robot video, and estimates that its captioning system could cost $0.72 per video hour. The lab compares that modeled figure with a measured GPT-6 Astra API cost of $88.49 per hour, calling the difference approximately 125 times; the comparison is not a like-for-like measured production bill, and Flora’s license restricts use to research or evaluation.
Flora comprises Flora-TempSeg, a 9-billion-parameter model that identifies action intervals; Flora-Captioner, a 27-billion-parameter model that writes an instruction for each interval; and Flora-Rewarder, a 27-billion-parameter model that rates an episode or its shorter steps. According to the model cards and paper cited by ThorstenMeyerAI.com, the systems are built on Alibaba’s Qwen3.5 models and were used as a data engine for Nvidia’s GR00T-2 robot model.
The published cost comparison uses 50 episodes totaling 4,483 seconds, or about 75 minutes, from Nvidia’s DenseCaptionBench. GPT-6 Astra’s $88.49-per-hour figure is based on reported API use at public rates dated October 4, 2026; 106 of 7,062 attempts lacked usage reports, so the cited total covers known usage. Flora’s $0.72 figure is instead a model: it assumes eight GB300 GPUs rented at $10 per GPU-hour and projected throughput. The paper gives a range of $0.51 to $1.26 per video hour, excluding storage and overhead.
The arithmetic behind the headline is $88.49 divided by $0.72, or about 122.5, rounded to approximately 125 times. The cost figure applies to captioning, not to every Flora model or the full cost of operating a data pipeline. The paper also states that its production GPU-hour denominator has not been verified. Nvidia says Flora can process up to 200,000 hours of video per day on 512 GB300 GPUs; that is a stated capacity, not an independently verified deployment result in the material reviewed.
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 Costs Could Expand Robot Training Data
Video labels help developers link a robot’s actions to what happened and whether each step succeeded. Lower labeling costs, if achieved in practice, could make it less expensive to sort and evaluate large video collections used to train or monitor robot systems. That could reduce one data-processing bottleneck for robotics research and development.
The estimate does not settle the broader economics of building robots. It excludes some operating costs, and collecting useful robot video, acquiring hardware and establishing safety remain separate challenges. For businesses, the license is a central limit: the cited terms allow research or evaluation only, not commercial deployment. Flora may offer a reference point for commercial alternatives, but the source material does not establish it as a free production tool for companies.
NVIDIA GB300 GPU server for AI model inference
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Three Models, Three Labeling Tasks
Robot training video needs more than a general description. Timestamped captions identify what action occurs and when; step-level process rewards mark actions as normal, suboptimal or failed; and outcome scores rate an entire attempt. Flora separates these functions across its segmenter, captioner and rewarder, rather than treating video labeling as one task.
According to the project information summarized by ThorstenMeyerAI.com, Flora was trained on fewer than 10,000 hours of human and robot video. The captioner supports a main camera or a main camera combined with wrist cameras. The models are distributed through Hugging Face repositories under the NVIDIA OneWay Noncommercial License. The source reports that derivatives inherit the noncommercial restriction and that Nvidia may use the models commercially.
The benchmark results are not uniformly in Flora’s favor. On the reported combined DenseCaptionBench score, Flora scored 57.1% against GPT-6 Astra’s 53.7%. Flora did better on human videos, action-boundary segmentation and partial-success cases. Astra performed better on robot-video caption accuracy and rubric scoring. Those results come mainly from Nvidia-run tests, and the source does not describe an independent evaluation of the cost estimate.
GPU cloud computing rental for AI video processing
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Production Costs and Quality Need Validation
The 125× comparison contrasts a measured API subtotal with a projected GPU budget, using an assumed rental price and modeled throughput. The Flora estimate excludes storage and overhead, and the paper says the GPU-hour denominator has not been verified. The source also notes the benchmark’s 50 episodes were selected by Nvidia; it is not clear whether the same cost and quality results would hold across larger or different video collections.
Quality is also task-dependent. Flora’s reported overall score was higher on the cited set, but GPT-6 Astra led on robot videos, caption accuracy and rubric scoring. The available material does not provide an independent replication of the full benchmark, nor does it establish how the models perform in a commercial pipeline. The reported 200,000-hours-per-day capacity is a vendor claim, while the stated noncommercial terms rule out treating the release as a general business offering.
robotics video annotation and labeling software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Tests and Deployment Costs
The next useful evidence would be independent benchmark replications and measured operating costs at production scale, including hardware rental, storage and other overhead. Tests across additional robot types, camera setups and longer video collections could show whether Flora’s reported advantages and costs generalize beyond Nvidia’s selected set.
For developers and companies, the immediate steps are to review the model cards, run permitted research or evaluation tests, and check the license before planning any use. The source material gives no date for a commercial license, broader release or independent production study, so those developments remain unannounced.
GPU workstation for large language model inference
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 three-model system from Nvidia’s GEAR robotics lab for segmenting actions in videos, writing captions for those segments and scoring robot performance.
Is the 125× cost reduction a measured result?
No. The comparison uses a modeled Flora cost of $0.72 per video hour and a measured GPT-6 Astra API usage subtotal of $88.49 per hour. Nvidia’s paper says the production GPU-hour denominator has not been verified.
Does the cost estimate cover all three Flora models?
No. The cited 125× comparison is for captioning. It does not include the Flora-Rewarder or establish the full cost of operating all three models in a production pipeline.
Can companies use Flora commercially?
The source says the models are released under the NVIDIA OneWay Noncommercial License, which permits research or evaluation use. The described terms do not permit commercial use, and derivatives inherit the restriction.
Did Flora outperform GPT-6 Astra on every test?
No. Flora scored higher on the reported combined benchmark and did better on human videos, segmentation and partial-success cases. GPT-6 Astra led on robot videos, caption accuracy and rubric scoring, according to the reported results.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
