🔍 Read the full analysis: Nvidia Flora And Robot Video Labels: Understanding The Savings And License 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, describing and scoring robot video. Nvidia estimates labeling costs at 72 cents per video hour, compared with $88.49 for GPT-6 Astra on a selected test set, but Flora’s figure is modeled and its license permits research and evaluation, not commercial use.
Nvidia’s GEAR robotics lab has released Flora, a set of three models that segment, describe and score video of robot actions, and estimates they can label video for $0.72 per hour. That figure is a modeled GPU cost, not a measured production bill; Nvidia’s comparison puts GPT-6 Astra at $88.49 per hour on a 50-episode test set. Flora’s license also limits use to research or evaluation, excluding commercial deployment.
The three models handle different labeling tasks. Flora-TempSeg, a 9-billion-parameter model, identifies the time intervals in which actions occur. The 27-billion-parameter Flora-Captioner writes an instruction for each interval, while the 27-billion-parameter Flora-Rewarder scores an overall episode, grades steps as normal, suboptimal or failure, or applies a user-provided rubric. The models are based on Alibaba’s Qwen3.5 models, according to the source report and Flora materials.
Nvidia’s cost comparison uses 50 episodes totaling 4,483 seconds—about 75 minutes—from its DenseCaptionBench. The reported GPT-6 Astra cost, $88.49 per video hour, is based on measured API usage at public rates dated October 4, 2026. Flora’s $0.72 figure is calculated from an assumed $10 per GPU-hour and modeled throughput using eight GB300 GPUs. The paper gives a range of $0.51 to $1.26 per hour and says the production GPU-hour denominator has not been verified. The quoted comparison concerns captioning, not the reward model.
On the same episodes, Flora scored 57.1% and GPT-6 Astra scored 53.7%, according to the report. Results varied across tasks: Flora performed better on the human-video subset and partial-success scoring, while Astra led on robot videos, caption accuracy and rubric scoring. Nvidia ran most of the tests, so the results are not an independent evaluation. The models are distributed under the NVIDIA OneWay Noncommercial License, which permits research and evaluation; derivatives inherit the restriction, according to the source.
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
Video labels help robotics teams identify what happened during a task and where a robot succeeded or failed. Automating that work could reduce one expense involved in preparing large training datasets. Nvidia says Flora was used as a data engine for its GR00T-2 robot models and estimates a capacity of up to 200,000 video hours a day on 512 GB300 GPUs. That is a company-reported capacity figure, not evidence that customers have achieved that throughput.
The cost comparison offers a potential benchmark, but its figures are not directly equivalent: one is a measured API cost and the other a modeled GPU budget with an assumed rental rate. Actual costs may depend on hardware availability, throughput, storage and operating overhead. The noncommercial license is a separate practical limit. Companies may study or evaluate Flora, but the stated terms do not grant permission to use it as a commercial labeling service or production pipeline.
Cheaper labeling would not by itself resolve other constraints on robot development. Data collection, hardware expense and safety remain challenges, and the reported test measures labeling performance on a limited set of episodes rather than real-world robot reliability.
NVIDIA GPU workstation for AI model inference
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Three Models for Robot Video
Robot-learning datasets can include first-person human demonstrations, gripper footage and recordings of more dexterous robots. As described in the Flora paper, training workflows use timestamped captions to identify actions, step-level rewards to mark progress or failure, and outcome scores for complete attempts. Those labels can help teams organize data and evaluate behavior.
Flora brings those tasks into one model family, but the benchmark results are mixed rather than a clear win across every category. The source report says Flora led on action segmentation in WGO-Bench, scoring 50.3 against Astra’s 46.0, and on the human subset of Nvidia’s DenseCaptionBench, scoring 37.4 against 29.7. On robot-video captioning and rubric scoring, Astra performed better. Flora’s model cards say it was trained on fewer than 10,000 hours of human and robot video. Its 27-billion-parameter models also require substantial computing resources; the report estimates roughly 55 GB for each model’s weights at 16-bit precision.
As an affiliate, we earn on qualifying purchases.
Costs and Results Still Need Testing
The real-world cost of running Flora is not established by the reported estimate. Nvidia’s calculation assumes a GPU-hour price and modeled throughput; the paper says its production GPU-hour denominator has not been verified, and the estimate excludes storage and overhead. The comparison also uses a fixed set of 50 episodes selected from Nvidia’s benchmark, so its results may not represent other datasets or workloads.
It is also unclear how Flora performs in independent evaluations or at sustained production scale. The provided source reports mixed benchmark results and says Nvidia conducted most tests. The license’s application to particular commercial arrangements should be checked against its full text; the source’s stated terms are research or evaluation only, with derivatives retaining the restriction.
robotics development kit with camera and sensors
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Tests and License Review
The next useful evidence would include independent benchmark results, measured operating costs across production workloads and clearer throughput data under real deployment conditions. Companies considering Flora will also need to review the NVIDIA OneWay Noncommercial License before using or adapting the models, particularly if their work has a commercial purpose. Until those details are established, the cost figure is best treated as Nvidia’s modeled estimate, not a guaranteed savings rate.
high speed NVMe SSD for AI video datasets
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Nvidia Flora?
Flora is a set of three models from Nvidia’s GEAR robotics lab for robot-video labeling: one segments action intervals, one writes captions, and one scores steps or full episodes.
Is Flora really 125 times cheaper than GPT-6 Astra?
Nvidia’s reported ratio is about 122.5, rounded to approximately 125. It compares a modeled Flora GPU cost of $0.72 per video hour with measured GPT-6 Astra API usage of $88.49 on the same selected test episodes. The figures use different costing methods, and Flora’s estimate is not a measured production bill.
Can a company use Flora commercially?
The source report says the models use the NVIDIA OneWay Noncommercial License, limited to research or evaluation, and that derivatives inherit the restriction. Companies should review the license text for their specific use; the stated terms do not offer a general commercial-use grant.
Does Flora outperform GPT-6 Astra on every test?
No. The reported scores favor Flora on some tasks, including human videos and partial-success scoring, while Astra leads on robot videos, caption accuracy and rubric scoring. Nvidia conducted most of the reported evaluations.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
