🔍 Read the full analysis: The License Question Behind Nvidia Flora’s Cheaper Robot Video Labeling on ThorstenMeyerAI.com
Get the latest gadgets 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 has released Flora, three models for segmenting, captioning and scoring robot-video data. Nvidia estimates captioning costs $0.72 per video hour, compared with $88.49 for GPT-6 Astra on its test, but the Flora figure is modeled rather than a measured production bill. The models use a noncommercial license, and benchmark results vary by task.
Nvidia’s GEAR robotics lab has released Flora, a set of three models that label video for robot training, and estimates its captioning pipeline costs $0.72 per hour of video, compared with $88.49 for GPT-6 Astra in the lab’s test. The headline cost is a model based on assumed GPU rental rates, not a measured production bill, and Flora’s license limits use to research or evaluation, excluding ordinary commercial deployment.
Flora is made up of three models built on Alibaba’s Qwen3.5: a 9-billion-parameter segmenter that identifies action boundaries, a 27-billion-parameter captioner that describes actions, and a 27-billion-parameter rewarder that grades robot performance. Nvidia says the tools were used as a data engine for its GR00T-2 robot models. The model cards say Flora was trained on fewer than 10,000 hours of human and robot video.
The cost comparison uses 50 episodes from Nvidia’s DenseCaptionBench, totaling 4,483 seconds, or about 75 minutes of video. Nvidia reports $88.49 per video hour for GPT-6 Astra at maximum reasoning effort, using public API rates dated October 4, 2026. Flora’s $0.72 figure is calculated using eight GB300 GPUs priced at an assumed $10 per GPU-hour and modeled throughput. The paper gives a range of $0.51 to $1.26 per hour, excluding storage and other overhead.
Nvidia reports Flora scoring 57.1% on the test’s quality measure, compared with 53.7% for Astra. The results are not uniformly in Flora’s favor: the paper says Flora performs better on human videos, action boundaries and partial-success cases, while Astra leads on robot videos, caption accuracy and rubric scoring. The cost claim 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-training systems rely on video labels that identify what happened, when an action began and whether a step succeeded. Producing those labels at scale can consume time and money, so a lower-cost tool could make it easier for robotics teams to process large datasets. Nvidia says Flora can label up to 200,000 hours of video per day on 512 GB300 GPUs, a company-reported capacity figure rather than an independently verified production result.
The economic comparison may be meaningful, but it should not be read as a like-for-like invoice comparison. Astra’s number is based on reported API usage, while Flora’s is a modeled GPU budget using assumed rental pricing and throughput. The evaluation also covers a relatively short, Nvidia-selected set of episodes. It offers evidence about those tests, not a guarantee of similar cost or quality across different datasets and operating conditions.
The license is a direct constraint for companies considering the models. Flora’s three components are released under the NVIDIA OneWay Noncommercial License, which permits research or evaluation use; derivatives inherit the restriction. That makes the models available for study and comparison, but the release does not amount to a free production labeling service for commercial robot development.
Products Worth Considering
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
Apple iPad Air (4th Generation, 10.9-inch, Wi-Fi, 256GB) – Space Gray
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
As an affiliate, we earn on qualifying purchases.
How Flora Labels Robot Video
Robot models learn from recordings of tasks such as placing objects, handling tools or folding fabric. A video can be made more useful for training when it is paired with timestamped descriptions of actions and judgments about whether individual steps succeeded. The Flora paper describes labels for action intervals, step-by-step process rewards and scores for an entire episode.
Flora’s segmenter returns time ranges for operations, and its captioner writes an English instruction for each range using surrounding video for context. The rewarder can score an episode from 1 to 5, classify moments as normal, suboptimal or failure, or apply a user-provided rubric. The model cards say the tools can work with a main camera and, in some setups, additional wrist cameras.
Nvidia’s paper says the system was developed by its GEAR lab, associated with Yuke Zhu and Linxi “Jim” Fan. The reported evaluation used Nvidia’s own DenseCaptionBench and compared Flora with several commercial models. The paper also cautions that the GPU-hour denominator for production has not been verified, a limitation that matters when interpreting the cost estimate.
“Research or evaluation purposes only.”
— NVIDIA OneWay Noncommercial License
Cost and Benchmark Questions
The $0.72 estimate depends on an assumed $10 per GPU-hour price and modeled throughput. Nvidia provides scenarios from $0.51 to $1.26, but says the production GPU-hour denominator has not been verified; storage and overhead are excluded. The comparison is also asymmetric: Astra’s figure comes from API usage, while Flora’s is estimated from GPU costs.
The test covers 50 episodes and roughly 75 minutes of video, and the source material does not establish whether results carry over to other robot types, camera setups, video lengths or operating conditions. Nvidia conducted most of the reported testing. The available information also does not provide independent confirmation of the daily throughput claim. Although the paper reports different strengths for Flora and Astra across tasks, the supplied results do not establish a single model as best for every labeling need.
The license’s practical boundaries for specific commercial arrangements may require legal review. The stated terms restrict use to research or evaluation and extend the restriction to derivatives; Nvidia may use its own models commercially. The source material does not explain what licensing options, if any, might later be offered to outside companies.
Commercial Access and Further Tests
The immediate next step for robotics teams is to review Flora’s model cards, paper and license before testing the models. Researchers can evaluate whether the tools suit their video data, but companies seeking production use will need to address the noncommercial restriction rather than assume the released weights can be deployed in a paid service.
Further evidence would help establish whether the cost estimate holds under measured production workloads, including GPU utilization, storage and operating overhead. Broader evaluation by teams outside Nvidia could also show how performance changes across datasets and robot tasks. Nvidia’s release describes the models and its own benchmarks, but the source material gives no timetable for independent studies, a commercial license or a measured production-cost update.
Key Questions
What is Nvidia Flora?
Flora is a group of three models for robot-video labeling: one segments actions by time, one captions those segments, and one scores robot performance.
Is Flora’s $0.72-per-hour cost measured?
No. Nvidia’s figure is a modeled estimate based on assumed GPU rental costs and throughput. The paper says the production GPU-hour denominator has not been verified, and the estimate excludes storage and overhead.
Can a company use Flora commercially?
The released models use the NVIDIA OneWay Noncommercial License, which states that use is for research or evaluation. The supplied material does not describe a commercial license for outside companies.
Did Flora outperform GPT-6 Astra?
On Nvidia’s reported test, Flora scored 57.1% and Astra 53.7%. Results varied by task: Flora led on some human-video and partial-success measures, while Astra performed better on robot videos, caption accuracy and rubric scoring.
Does the 125-times cost comparison cover all of Flora?
No. The comparison concerns captioning costs on a 50-episode test. It does not include Flora’s separate reward model, and it compares an estimated GPU budget with reported API usage.
Source: ThorstenMeyerAI.com
