🔍 Read the full analysis: Nvidia Flora: A 125× Cheaper Way To Label Robot Videos, With A Licensing Twist 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, describing and scoring robot-video actions. Nvidia estimates a video-labeling cost of $0.72 per hour versus $88.49 for GPT-6 Astra, but Flora’s cost is modeled, the benchmark is limited, and its license permits research or evaluation rather than commercial use.
Nvidia’s GEAR robotics lab has released Flora, three models designed to label robot-video data, and estimates that running them costs $0.72 per hour of video compared with $88.49 using GPT-6 Astra. The headline cost advantage is not a measured production bill: Nvidia modeled Flora’s GPU costs, and the models’ OneWay Noncommercial License limits use to research or evaluation, restricting commercial deployment.
Flora comprises a 9-billion-parameter temporal segmenter that finds action boundaries, a 27-billion-parameter captioner that describes each segment, and a 27-billion-parameter reward model that scores steps or full episodes. Nvidia says the suite was used as a data engine for its GR00T-2 robot models and was trained on fewer than 10,000 hours of robot and human video. The project’s models and inference code are available through Hugging Face.
Nvidia’s cost comparison uses 50 episodes from its DenseCaptionBench, totaling 4,483 seconds, or about 75 minutes. The company reports an Astra API cost of $88.49 per video hour, based on public rates and known usage. Flora’s $0.72 figure is a modeled estimate using eight GB300 GPUs at an assumed $10 per GPU-hour. Nvidia gives a range of $0.51 to $1.26 and says the estimate excludes storage and overhead; its paper states that the production GPU-hour denominator has not been verified. The ratio of the two headline figures is about 122.5 to one, rounded by Nvidia to approximately 125×.
The benchmark results are mixed. On the same test episodes, Nvidia reports Flora at 57.1% quality, ahead of Astra at 53.7%, Gemini 3.8 Flash at 38.3%, and Claude Opus 5 at 30.4%. Flora also scores better on the human-video subset and partial-success evaluation cited in the source material. Astra performs better on robot videos, caption accuracy and rubric scoring. The evaluations were largely conducted by Nvidia, and the cost comparison covers 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, Limited Use
Video annotations tell robot-learning systems what action occurred and when, and can mark failed or incomplete steps. If those labels can be generated at lower cost, researchers may be able to process more training footage or test more approaches to robot learning. Nvidia says Flora supports labeling at a scale of up to 200,000 hours of video per day on 512 GB300 GPUs; that is a company-reported capacity claim, not an independent production measurement.
The licensing terms complicate the practical value for businesses. All three models use the NVIDIA OneWay Noncommercial License, which permits research or evaluation and passes the same restriction to derivatives. The source says Nvidia may use the models commercially, but companies cannot treat the release as a free production labeling service under the stated terms. For commercial teams, Flora may serve as a research tool or comparison point, while production use would require a different license or another approach.
Cheaper annotation addresses only one part of robot development. It does not remove the expense of collecting suitable video, building and maintaining robots, or managing safety when systems act in homes and workplaces. The claimed saving is relevant to data-processing budgets, but it does not establish that capable, affordable robots will arrive sooner.
Popular right now
Philips 4400 Series Fully Automatic Espresso Machine with LatteGo Milk System
TCL 85 Designer Series QD-Mini LED TV (Model 85A400U) with Google Smart TV
Breville Bambino Espresso Machine, Brushed Stainless Steel
As an affiliate, we earn on qualifying purchases.
Three Models for Video Labels
Robot-learning datasets need more than a video file. According to Nvidia’s paper, timestamped captions link instructions to the relevant moments; process rewards grade short steps as normal, suboptimal or failed; and outcome rewards score an entire attempt. These labels can help filter training data and assess how a robot performed. Flora divides those tasks among its segmenter, captioner and rewarder.
The models are built on Alibaba’s Qwen3.5 family. Flora-TempSeg is an adapter on Qwen3.5-9B; the captioner and rewarder are based on Qwen3.5-27B. The captioner can take a single camera view or a main view alongside wrist cameras. The rewarder can grade half-second intervals or score an episode against a user-supplied rubric. Nvidia lists an H100 running Linux as test hardware; the source estimates that each 27B model requires roughly 55 GB of memory for 16-bit weights, making them data-center-scale tools rather than ordinary gaming-PC software.
The approximately 125× claim should be read as a particular comparison, not a universal operating-cost result. Astra’s figure is based on API charges for a selected 50-episode set, while Flora’s is a projected GPU budget at an assumed rental rate and modeled throughput. Nvidia’s paper notes that 106 of 7,062 attempts reported no usage, so the Astra amount is a known-usage subtotal. The source also reports that The Decoder characterized the estimates as projections rather than production measurements.
“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 Quality Still Need Testing
It is not yet clear what Flora will cost in a sustained production deployment. Nvidia’s estimate depends on an assumed $10 GPU-hour price and modeled throughput, and excludes storage and overhead; the paper says the production GPU-hour denominator has not been verified. Actual costs would depend on hardware access, utilization, engineering and the quantity and type of footage processed.
The benchmark does not settle how Flora performs across the broader range of robot tasks or camera setups. The reported test contains 50 episodes selected from Nvidia’s own benchmark, and the source says Nvidia ran most of the tests. Flora’s results vary by category, with Astra retaining advantages on robot videos, caption accuracy and rubric scoring. Independent replications and further testing on varied datasets would help establish how well the results generalize.
The license’s application to specific commercial research, derivative models and mixed research-production workflows may require legal interpretation. The supplied material establishes the noncommercial restriction, but does not provide a separate commercial license or clarify whether Nvidia plans to offer one. The 200,000-hours-per-day figure is also Nvidia’s stated capacity, not an independently verified measure of routine output.
Independent Tests and Licensing Terms
The next useful evidence will be independent evaluations on additional robot and human-video datasets, along with clearer reporting of real-world GPU throughput and full operating costs. Comparing model quality and cost across the same footage, hardware and accounting assumptions would help distinguish an achievable saving from a modeled scenario.
Commercial developers will also need to check the license before incorporating Flora into products or production workflows. Nvidia’s model cards and paper provide the stated terms and technical details; whether the company will offer separate commercial rights is not specified in the material available as of October 25, 2026. Until then, the release offers researchers a new labeling suite, while its cost advantage and commercial applicability remain qualified.
Key Questions
What is Nvidia Flora?
Flora is a three-model suite from Nvidia’s GEAR robotics lab. Its models identify action boundaries in video, write captions for those actions and score robot performance.
Is the 125× cost saving measured?
No. Nvidia compares Astra’s reported API usage cost of $88.49 per video hour with a modeled Flora GPU cost of $0.72. The Flora estimate assumes $10 per GPU-hour, excludes storage and overhead, and has not been verified as a production cost.
Can a company use Flora commercially?
The supplied source says the models use Nvidia’s OneWay Noncommercial License, for research or evaluation only, with the restriction applying to derivatives. It does not identify a separate commercial license.
Does Flora outperform GPT-6 Astra on every test?
No. Nvidia reports Flora ahead on its overall score for the 50-episode test and on some human-video and partial-success measures. Astra leads on robot videos, caption accuracy and rubric scoring in the comparisons described in the source.
What hardware does Flora require?
The source lists an Nvidia H100 running Linux as test hardware and estimates that each 27-billion-parameter model needs about 55 GB for 16-bit weights. That points to data-center-class hardware rather than a typical gaming PC.
Source: ThorstenMeyerAI.com
