🔍 Read the full analysis: Flora’s 125× Cost Savings For Robot Video Labeling Come With License Terms 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 72 cents per video hour, about 122.5 times below the $88.49 API cost it reports for GPT-6 Astra on its test set; Flora’s figure is modeled, and the models’ license permits research or evaluation, not commercial use.
Nvidia’s GEAR robotics lab has released Flora, a set of three models designed to label video used to train robots, and estimates that its captioning system can process video for $0.72 per hour—far below the $88.49 per hour Nvidia reports for GPT-6 Astra on the same test episodes. The comparison is not like-for-like in how costs were calculated: Flora’s price is a modeled GPU estimate, and its OneWay Noncommercial License bars commercial use.
Flora combines three models built on Alibaba’s Qwen3.5 family. A 9-billion-parameter segmenter identifies action intervals in a video; a 27-billion-parameter captioner describes those intervals in English; and a second 27-billion-parameter model scores robot performance. Nvidia says the tools were used as a data engine for its GR00T-2 robot models. The project paper and model cards describe training on fewer than 10,000 hours of human and robot video.
Nvidia’s cost comparison uses 50 episodes totaling 4,483 seconds, or about 75 minutes, from its DenseCaptionBench. It reports GPT-6 Astra API usage of $88.49 per video hour, using public rates dated October 4, 2026. Flora’s $0.72 estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled processing throughput; Nvidia gives a range of $0.51 to $1.26, excluding storage and overhead. The $88.49 divided by $0.72 is about 122.5, rounded in the claim to approximately 125 times.
The paper reports an overall score of 57.1% for Flora on those episodes, compared with 53.7% for Astra, 38.3% for Gemini 3.8 Flash and 30.4% for Claude Opus 5. Results vary by task: Flora performs better on human videos, action boundaries and some failed or partially successful actions, while Astra leads on robot-video caption accuracy and rubric scoring. Nvidia conducted most of the reported tests itself. The cost ratio concerns captioning and does not include the 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, Limited Access
Video annotation is a significant expense in building robot-training datasets: someone or something must identify what happens, when an action starts and ends, and whether it succeeds. If Flora’s estimated cost can be achieved at scale, lower labeling costs could make it easier for robotics teams to process larger video collections. That could reduce one obstacle to training robots on a wider range of tasks, though it does not resolve the separate costs of collecting useful footage, building hardware or testing systems safely.
The license changes the immediate value for companies. Flora is not a free commercial labeling service: the supplied models are licensed for research or evaluation, and the source material says derivatives inherit the same restriction. Nvidia may use the models commercially, but other organizations cannot assume that releasing downloadable weights gives them permission to run Flora in a paid production workflow. For businesses, the work may serve as a technical reference or a research tool rather than a ready-to-use production pipeline.
Popular right now

Ninja Luxe Café Premier 3-in-1 Espresso Machine, Drip Coffee & Cold Brew

Ninja Luxe Café Mini Plus Espresso and Drip Coffee Machine with AutoFroth, Stainless Steel, ES351

De'Longhi Magnifica Duo Fully Automatic Espresso Machine with Milk Frother, Silver
As an affiliate, we earn on qualifying purchases.
Three Models for Robot Video
Robot-learning systems use video of both people and machines performing tasks. The labeling described in the Flora paper includes timestamped captions tied to intervals, step-by-step process grades such as normal, suboptimal or failure, and scores for an entire attempt. These labels can help researchers organize training material and identify where a robot’s behavior went wrong. Flora splits those jobs across its segmenter, captioner and rewarder rather than treating all labels as one output.
The models are available through Hugging Face, with inference code in their repositories. The source material says testing used an Nvidia H100 on Linux and estimates that each 27-billion-parameter model requires roughly 55 GB for weights at 16-bit precision. That points to substantial computing requirements, not ordinary consumer hardware. Nvidia also says its broader system can label up to 200,000 hours of video per day on 512 GB300 GPUs; that is a company-reported capacity claim, not a result established by the 50-episode cost test.
“expensive at the scale of modern robot datasets”
— Nvidia, describing the cost of frontier models in the Flora paper
Cost and Quality Need Validation
The headline cost advantage remains an estimate, not a controlled comparison of two production systems’ measured bills. Nvidia compared measured API usage for Astra with Flora’s modeled GPU budget, using a small test set selected from Nvidia’s own benchmark. The paper also says 106 of 7,062 Astra attempts reported no usage, making its figure a known-usage subtotal. Flora’s stated range depends on assumed GPU pricing and throughput, and excludes storage and overhead. It is not clear from the supplied material whether independent teams have reproduced the throughput or cost estimate in routine production.
Performance is also task-dependent. Flora’s higher overall score on the selected episodes does not mean it leads on every benchmark or type of footage; the source says Astra remains ahead on robot videos, caption accuracy and rubric scoring. The supplied details do not establish how the models perform across broader datasets, different camera setups or tasks outside the reported tests. Nor do they establish commercial rights beyond the stated noncommercial license. Nvidia’s capacity figure and the cost estimates should be read as company-reported claims, not independently verified outcomes.
Independent Tests and License Questions
The immediate next step for researchers is to examine the paper, model cards and license, then test Flora against their own footage and hardware budgets. Such tests could show whether the modeled throughput and quality hold beyond Nvidia’s benchmark episodes. Organizations considering deployment will also need to review the OneWay Noncommercial License and obtain appropriate permission before any commercial use; the supplied material does not describe a commercial license for other users.
Further reporting will depend on whether independent evaluations confirm the cost and quality results, whether Nvidia updates its model documentation or license terms, and whether the models see wider use in research. Until then, the reported 125-times figure is best understood as a rounded comparison of a modeled Flora cost with reported Astra API spending, not a guarantee of savings for a company’s production workload.
Key Questions
What is Flora?
Flora is a set of three video models from Nvidia’s GEAR robotics lab: one segments actions by time, one captions them, and one scores robot performance.
Is the 125-times cost saving measured?
No. Nvidia’s $0.72 per video hour for Flora is modeled using assumed GPU rental pricing and throughput. The $88.49 Astra figure is reported API usage on the same 50 episodes. Nvidia also notes that the production GPU-hour denominator has not been verified.
Can a company use Flora commercially?
The supplied models use Nvidia’s OneWay Noncommercial License, which permits research or evaluation and bars commercial use. The source says derivatives inherit that restriction; companies should review the license rather than treating model availability as commercial permission.
Does Flora outperform GPT-6 Astra on every test?
No. Flora scored higher overall on Nvidia’s 50-episode test and did better on some human-video, action-boundary and partial-success tasks. The source says Astra led on robot videos, caption accuracy and rubric scoring.
Source: ThorstenMeyerAI.com
