🔍 Read the full analysis: Robot Video Labeling Gets Cheaper With Nvidia Flora—but The License Matters on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
Nvidia’s GEAR lab released Flora, three models for segmenting, captioning and scoring robot videos. Nvidia estimates captioning costs $0.72 per video hour, versus $88.49 for GPT-6 Astra on a selected test set, but Flora’s figure is modeled, quality results are mixed, and its license restricts use to research or evaluation.
Flora includes a 9-billion-parameter temporal segmenter that identifies action intervals, a 27-billion-parameter captioner that writes instructions for those intervals, and a 27-billion-parameter rewarder that scores a full episode or grades actions at half-second intervals. The models are based on Alibaba’s Qwen3.5 family. Nvidia’s model cards say Flora was used as a data engine for its GR00T-2 robotics models and was trained on fewer than 10,000 hours of human and robot video.
The cost estimate comes from a test of 50 episodes totaling about 75 minutes, drawn from Nvidia’s DenseCaptionBench. GPT-6 Astra’s $88.49-per-video-hour cost reflects reported API usage at public rates; Nvidia’s Flora estimate of $0.72 assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput. The paper gives Flora scenarios from $0.51 to $1.26 per video hour and says the production GPU-hour denominator has not been verified. The estimate excludes storage and overhead, and the reported comparison concerns captioning, not the reward model.
Nvidia reported an aggregate quality score of 57.1% for Flora, versus 53.7% for Astra, on the same set. Its results vary by task: the paper reports Flora ahead on human-video captions, action boundaries and some partial-success judgments, while Astra scores better on robot-video captioning, caption accuracy and rubric scoring. The testing was largely conducted by Nvidia, and the episodes came from its own benchmark. The reported approximately 125-fold cost gap is the rounded ratio of $88.49 to $0.72, not a measured production-cost comparison.
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 Access
Robot training depends on turning footage into usable descriptions: what an arm did, when an action occurred and whether it succeeded. If labeling can be done for less, developers may be able to process more footage or reduce one expense in building training datasets. Nvidia says its system can label up to 200,000 hours of video per day using 512 GB300 GPUs; that is a vendor-reported capacity figure, not an independent production measurement.
For companies, however, the license changes the practical value of the release. All three models are offered under the NVIDIA OneWay Noncommercial License, which permits research or evaluation and carries the same restriction to derivatives. The source material says Nvidia itself may use the models commercially. Businesses therefore cannot treat the public release as a ready-made, free commercial labeling service. The results may still provide a technical reference point, but companies would need to resolve licensing or use another permitted approach before deployment.
Even if labeling becomes cheaper, it does not resolve other constraints on robotics: collecting suitable robot data, ensuring safe behavior and paying for hardware remain separate challenges. Flora addresses a data-processing task, not the full cost or difficulty of developing and operating robots.
NVIDIA GB300 GPU server workstation
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Three Models for Video Labels
Robot-learning datasets contain videos of tasks such as placing objects, passing tools or folding fabric. Training systems may need timestamped captions to locate and describe actions, process rewards to identify good or failed steps, and outcome scores for an entire attempt. Flora divides those functions among its segmenter, captioner and rewarder rather than treating video labeling as one task.
The models are available through Hugging Face repositories, with inference code included there, according to the source material. Testing used an Nvidia H100 running Linux. The two 27-billion-parameter models require substantial memory: the source estimates roughly 55 GB for weights at 16-bit precision, putting them in data-center-class territory rather than within the reach of many consumer computers.
Nvidia’s cost calculation uses a fixed 50-episode test set spanning human hands and different robot hands. The API baseline uses Astra at maximum reasoning effort and October 4, 2026, public rates. The difference between a measured usage-based API bill and a modeled GPU budget is central to interpreting the headline savings.
Cost and Quality Still Need Testing
The $0.72 figure is a modeled estimate, not an observed bill from a production deployment. Nvidia’s paper states that the production GPU-hour denominator has not been verified; actual costs could vary with hardware rates, utilization, workload and overhead. The comparison also pairs a modeled GPU budget with Astra’s reported API usage and is based on roughly 75 minutes of video selected from Nvidia’s benchmark.
Quality results do not establish that Flora is better across all robot-video labeling tasks. Nvidia’s reported aggregate score favors Flora on the selected episodes, while Astra leads on several robot-video and scoring measures. Independent evaluations, performance on different datasets and real-world production costs remain unclear. The source also does not establish whether Nvidia will offer a commercial license or a separate paid service for businesses.
Independent Tests and Licensing Terms
The next useful evidence would be independent evaluations on broader video collections, alongside production cost measurements that include hardware utilization, storage and operational overhead. Those tests could show whether the estimated savings and mixed quality results hold beyond Nvidia’s chosen episodes.
Companies considering Flora will also need to review the license before using the models or derivatives in commercial workflows. Until Nvidia changes or clarifies those terms, the release is available for research and evaluation, while its suitability as a business production tool remains limited by the stated noncommercial restriction.
Key Questions
What is Nvidia Flora?
Flora is a three-model suite for robot-video data: a segmenter locates action intervals, a captioner describes them, and a rewarder scores actions or full episodes.
Is the $0.72-per-hour cost a measured production price?
No. Nvidia’s $0.72 estimate is modeled using assumed GPU rental costs and throughput. Its paper says the production GPU-hour denominator has not been verified and excludes storage and overhead.
Does Flora cost 125 times less than GPT-6 Astra in every use?
No. The approximately 125-fold figure is the rounded ratio between $88.49 per video hour for Astra API usage and Flora’s modeled $0.72 estimate on a selected 50-episode captioning test. It is not a general measured production comparison.
Can a company use Flora commercially?
The models are released under the NVIDIA OneWay Noncommercial License, which the source describes as allowing research or evaluation only. Businesses should review the license and cannot assume the models are cleared for commercial deployment.
Does Flora outperform Astra on every benchmark?
No. Nvidia’s results are mixed by task: Flora leads on some human-video, segmentation and partial-success tests, while Astra scores better on some robot-video captions and scoring measures.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
