🔍 Read the full analysis: Nvidia Flora’s Robot Video Labeling Deal: 125× Lower Cost, License Caveat on ThorstenMeyerAI.com
Get smart everyday buys 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 Flora’s captioning cost at $0.72 per video hour, compared with $88.49 for GPT-6 Astra, but the comparison pairs a modeled GPU cost with measured API use, and Flora’s license restricts use to research or evaluation.
Nvidia’s GEAR robotics lab has released Flora, a set of three models for labeling video used to train and evaluate robots, and estimates its captioning costs at $0.72 per video hour. That is about 122.5 times below the $88.49-per-hour GPT-6 Astra figure in Nvidia’s comparison, but the company’s Flora cost is modeled rather than a measured production bill, and the models are licensed for research or evaluation only.
Flora comprises a 9-billion-parameter video segmenter, which marks where actions begin and end; a 27-billion-parameter captioner, which writes an instruction for each segment; and a 27-billion-parameter reward model, which scores a whole attempt or grades steps as normal, suboptimal or failures. The models are based on Alibaba’s Qwen3.5 family. Nvidia says they were used as a data engine for its GR00T-2 robot models.
The cost comparison in the Flora paper uses 50 episodes totaling 4,483 seconds—about 75 minutes—from Nvidia’s DenseCaptionBench. GPT-6 Astra’s $88.49 figure reflects reported API usage at public rates cited for October 4, 2026. Flora’s $0.72 figure is an estimate based on eight GB300 GPUs, an assumed price of $10 per GPU-hour and modeled throughput. The estimate ranges from $0.51 to $1.26 per video hour and excludes storage and other overhead. Nvidia’s paper says the production GPU-hour denominator has not been verified.
The headline ratio applies to captioning; the reward model is not included in that cost calculation. On the same episodes, Nvidia reports an overall score of 57.1% for Flora and 53.7% for Astra. The paper’s results are not uniformly in Flora’s favor: it reports Flora ahead on human-video captioning, action boundaries and partial-success evaluation, while Astra scores higher on robot-video captioning, caption accuracy and rubric scoring. Most tests were run by Nvidia, so independent replication would help establish how well the results generalize.
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
Cheaper Labels, With Limits
Video labels help robot developers identify what happened in a recording, connect instructions to specific moments and detect where a task went wrong. Producing those annotations at scale can be expensive, so lower-cost tools could reduce one part of the data-preparation burden for robotics research and development. Nvidia says Flora can process up to 200,000 hours of video per day using 512 GB300 GPUs; that is a company-reported capacity figure, not a guarantee of throughput or cost for other users.
The practical effect for businesses is less straightforward. Flora’s noncommercial license prevents companies from treating the released models as a free production service. Developers can examine and evaluate them under the stated terms, but commercial deployment requires a separate permission or another suitable license. The announced cost advantage is consequently a benchmark claim and research opportunity, not proof that a company can label its production data at the same price.
Even if labeling becomes cheaper, it does not remove other barriers to robotics. The source material identifies the collection of robot data, hardware costs and safety as continuing challenges. The reported comparison also concerns a selected test set and modeled GPU economics, rather than a broad measurement of operating costs across real-world datasets.
NVIDIA GB300 GPU server workstation
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Three Models for Robot Video
Robot training recordings can show a gripper placing a mug, hands passing an object or a person folding fabric. For a model to use those recordings, developers may need timestamped action descriptions, scores for individual steps and an assessment of the whole attempt. Flora divides those functions among three models rather than offering a single labeling tool.
According to the model cards described in the source material, the captioner can use one camera view or combine a main view with wrist-camera footage. The rewarder can give an episode an overall score from one to five, grade steps at half-second intervals or apply a user-provided rubric. Nvidia says Flora was trained on fewer than 10,000 hours of robot and human video. The inference code is included in the models’ Hugging Face repositories; the stated test setup uses an Nvidia H100 running Linux.
The paper’s benchmark results provide a mixed comparison rather than a universal ranking. Flora’s reported strengths include action segmentation and human videos, while the source account says Astra retains advantages on robot videos and some caption and rubric measures. The comparison should be read as results on the specified benchmarks and episodes, not as evidence that either system will lead on every robotics dataset.
“The production GPU-hour denominator has not been verified.”
— Nvidia, in the Flora paper
GPU server for AI video processing and model inference
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Cost and Commercial Use Questions
The Flora cost estimate depends on assumed GPU pricing and modeled throughput; Nvidia says the production GPU-hour denominator has not been verified. The comparison also sets that estimate against API usage for a limited, vendor-selected set of 50 episodes, and the source notes that 106 of 7,062 Astra attempts reported no usage, making the Astra amount a known-usage subtotal. The exact cost advantage in a production workflow is therefore not established by the benchmark.
It is also unclear from the available material how Flora’s performance would change across larger or more varied datasets, how much setup and storage add to operating costs, or whether outside teams can reproduce the reported scores. The license states that use is limited to research or evaluation and that derivatives inherit the restriction; the source material says Nvidia may use the models commercially. The terms for any separate commercial permission are not specified here.
robotics development kit with camera and sensors
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Replication and License Terms
The next useful test is independent evaluation on additional robot and human-video datasets, including measurement of actual GPU hours, storage and operating overhead. That would show whether the modeled price and reported quality hold beyond the 50-episode benchmark. Nvidia has released the models and inference code through Hugging Face, making the stated research and evaluation use possible under the license.
Organizations considering commercial deployment will need to review the license and determine whether they can obtain separate permission or must use another tool. Further details on commercial terms, broader production costs and performance at different scales have not been provided in the source material. Until those points are clarified, the release offers a research benchmark and a potential labeling approach, not a confirmed low-cost commercial pipeline.
high-capacity NVMe SSD for AI video datasets
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 suite of three models for robot-video labeling: a segmenter that identifies action intervals, a captioner that describes them and a rewarder that scores steps or whole attempts.
Is the $0.72-per-hour figure a measured production cost?
No. Nvidia’s figure is a modeled estimate based on assumed GPU pricing and throughput. The paper says the production GPU-hour denominator has not been verified, and the estimate excludes storage and other overhead.
What does the comparison with GPT-6 Astra show?
For captioning the paper’s 50 test episodes, Nvidia reports $0.72 per video hour for Flora and $88.49 for Astra. The ratio is about 122.5 to one, but it compares modeled GPU costs with reported API usage and does not establish the cost difference for every production workload.
Can a company use Flora commercially?
The stated NVIDIA OneWay Noncommercial License limits use to research or evaluation, and derivatives inherit that restriction. The source material says Nvidia may use the models commercially; it does not specify terms for other companies to obtain commercial permission.
Does Flora outperform Astra on every test?
No. Nvidia’s reported results show Flora ahead on some measures, including human-video captioning and action segmentation, while Astra leads on some robot-video captioning and scoring measures. The results cover the paper’s benchmarks, and independent replication remains useful.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
