🔍 Read the full analysis: Could Nvidia Flora Cut Robot Video Labeling Costs By 125×? Mind The License on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Nvidia’s GEAR lab released Flora, a set of three models for segmenting, captioning and scoring robot videos. Nvidia estimates its captioning cost at $0.72 per video hour versus $88.49 for GPT-6 Astra, but the 122.5-to-1 comparison combines a modeled GPU cost with measured API spending, and Flora’s license restricts use to research or evaluation.
Nvidia’s GEAR robotics lab has released Flora, three models designed to label robot-training videos, and estimates that its captioning system costs $0.72 per video hour compared with $88.49 for GPT-6 Astra. The headline comparison is not a like-for-like production bill: Nvidia’s figure is a modeled GPU cost, and the models’ noncommercial license permits research or evaluation, not ordinary business use.
The Flora set includes a 9-billion-parameter segmentation model that marks where actions begin and end, a 27-billion-parameter captioner that describes each segment, and a 27-billion-parameter rewarder that scores robot performance. The models are built on Alibaba’s Qwen3.5 models and are available through Hugging Face, according to the source report. Nvidia says Flora supplied labels for its GR00T-2 robotics work.
The reported cost comparison comes from tests on 50 video episodes totaling about 75 minutes. Nvidia’s paper puts GPT-6 Astra’s API cost at $88.49 per video hour and estimates Flora at $0.72, producing a ratio of 122.5, rounded to about 125 times. Flora’s estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput; the paper says the production GPU-hour denominator has not been verified. The estimate excludes storage and overhead, and Nvidia gives a range of $0.51 to $1.26 per video hour.
On the same episodes, the report says Flora scored 57.1% and GPT-6 Astra 53.7%. That result does not establish that Flora is better across tasks: the paper’s broader benchmark results are mixed, with Flora stronger on human videos, action boundaries and some partial-success assessments, while Astra leads on robot videos, caption accuracy and rubric scoring. The 125-times cost figure applies to captioning, not the separate 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, With Limits
Robot-learning systems need video labels to identify actions, describe what happened and judge whether a step succeeded. Automating that work could reduce one expense in building training datasets, especially when developers process large collections of human and robot footage. Nvidia says Flora can label up to 200,000 hours of video per day on 512 GB300 GPUs; that is a company-stated capacity, not an independently verified production result.
For companies evaluating the technology, however, price and access are separate issues. The license permits use for research or evaluation purposes only, and derivatives inherit the restriction, according to the report. That means businesses cannot treat the models as a free production labeling service under the stated terms. The release can still offer researchers a comparison point and a technical reference, but commercial deployment would require different permission or licensing terms.
Cheaper labeling would address only one part of robotics development. The report notes that collecting useful robot data, building reliable hardware and making systems safe remain unresolved costs and challenges. A low estimated labeling price alone does not show that a robot can perform tasks reliably outside a test set.
Popular right now
TCL 85 Designer Series QD-Mini LED TV (Model 85A400U) with Google Smart TV
De'Longhi Dinamica Plus Fully Automatic Espresso Machine
KitchenAid Artisan 5-Qt Tilt-Head Stand Mixer, Cardamom
As an affiliate, we earn on qualifying purchases.
Three Models for Video Labels
Robot models learn from recordings such as a gripper placing a mug or a person folding a towel. According to the Flora paper as described by ThorstenMeyerAI.com, developers use timestamped captions to connect instructions to moments in a recording, process rewards to grade individual steps, and outcome rewards to score an entire attempt. Those labels help organize training data and identify where a robot’s behavior went wrong.
Flora-TempSeg identifies intervals for separate operations. Flora-Captioner writes an English instruction for each interval and can use a main camera plus wrist-camera footage. Flora-Rewarder can assign an overall score from 1 to 5, grade actions at half-second intervals, or apply a user-provided rubric. The model cards reportedly say the models were trained on fewer than 10,000 hours of human and robot video.
The paper’s test used Nvidia’s DenseCaptionBench episodes: 20 showing human hands, 15 with two-finger grippers and 15 with dexterous robot hands. The cost comparison uses public API prices for GPT-6 Astra from October 4, 2026, against a modeled GPU budget for Flora. The report also notes that 106 of 7,062 Astra attempts had no usage reported, making the API figure a known-usage subtotal.
“The cost figure is a modeled estimate, the quality picture is mixed, and ‘open’ here means free for research, not for business.”
— Thorsten Meyer, ThorstenMeyerAI.com
Cost and Performance Still Unsettled
The reported ratio compares a modeled GPU budget with measured API usage, rather than two production deployments measured on the same accounting basis. The test also covers about 75 minutes of video selected for Nvidia’s benchmark, so it does not establish costs or results across other datasets, video lengths, hardware setups or operating conditions. The paper’s cost scenarios vary from $0.51 to $1.26 per hour and omit storage and overhead.
Performance is also task-dependent. The source report says Flora performed better on some human-video, segmentation and partial-success tests, while Astra performed better on robot-video captioning and rubric-based scoring. Most of the tests were run by Nvidia, according to the report. Independent replication, production-scale costs and how the models perform on a wider variety of footage remain unclear.
The license’s practical terms matter for potential adopters. The report identifies the NVIDIA OneWay Noncommercial License and says derivatives carry the same restriction. It does not establish whether Nvidia will offer a separate commercial license, on what terms, or whether the terms may change.
Replication and Licensing Will Matter
The next useful test is whether independent researchers can reproduce the benchmark results and cost assumptions on other video collections and hardware. Real-world cost comparisons would need to include GPU utilization, storage, setup and operating overhead, as well as the full range of labeling tasks rather than captioning alone.
Organizations considering Flora for a commercial pipeline will need to review the license and seek authorization or alternative terms before deployment. Until costs are verified at production scale and commercial rights are clarified, Nvidia’s reported ratio is best treated as a benchmark estimate, not a guaranteed operating saving.
Key Questions
What is Nvidia Flora?
Flora is a set of three models for robot-video labeling: one segments actions, one writes captions, and one scores performance.
Is Flora really 125 times cheaper than GPT-6 Astra?
Nvidia’s paper reports a ratio of about 122.5 to 1, rounded to 125. The comparison uses a modeled $0.72-per-hour GPU cost for Flora and $88.49 per hour of measured GPT-6 Astra API usage, so it is not a direct comparison of two verified production bills.
Can companies use Flora commercially?
According to the source report, the models use the NVIDIA OneWay Noncommercial License, which permits research or evaluation purposes only. Companies should not assume the models can be used in a commercial production pipeline under those terms.
Does Flora outperform GPT-6 Astra?
Results vary by task. The report says Flora scored 57.1% versus Astra’s 53.7% on the 50-episode comparison, but Astra led on robot-video captioning and rubric scoring, while Flora did better on some human-video, segmentation and partial-success tests.
What does the cost estimate leave out?
The Flora estimate assumes eight GB300 GPUs at $10 per GPU-hour and modeled throughput. It excludes storage and overhead, and Nvidia’s paper says the production GPU-hour denominator has not been verified.
Source: ThorstenMeyerAI.com
