🔍 Read the full analysis: Nvidia’s Flora Makes Robot Video Labeling 125× Cheaper. The Catch Is In The 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 robotics lab has released Flora, three video models for segmenting, captioning and scoring robot actions. Nvidia estimates Flora’s labeling cost at 72 cents per video hour, about 122.5 times below a GPT-6 Astra API figure, but the comparison pairs modeled GPU costs with measured API charges. Flora’s license restricts use to research or evaluation, and benchmark results vary by task.
Nvidia’s GEAR robotics lab has released Flora, a set of three models designed to mark actions, write captions and score performance in robot videos. Nvidia estimates the system can label video for $0.72 per hour, compared with $88.49 for GPT-6 Astra on the same test episodes; the figures are not like-for-like measurements, because Flora’s cost is modeled while Astra’s is based on API usage.
Flora includes a 9-billion-parameter segmenter that identifies action intervals, a 27-billion-parameter captioner that describes each interval, and a 27-billion-parameter reward model that evaluates a robot’s progress. The models are built on Alibaba’s Qwen3.5 family. Nvidia says Flora was used as a data engine for its GR00T-2 robot models and was trained on fewer than 10,000 hours of human and robot video.
The cost comparison in the paper uses 50 episodes totaling 4,483 seconds, or about 75 minutes, from Nvidia’s DenseCaptionBench. GPT-6 Astra’s $88.49 per video hour is calculated from reported API usage at public rates dated October 4, 2026. Flora’s $0.72 figure is a projection based on eight GB300 GPUs rented at an assumed $10 per GPU-hour and modeled throughput. Nvidia gives a range of $0.51 to $1.26 per hour, excluding storage and other overhead.
The arithmetic yields a ratio of about 122.5 to one, rounded by Nvidia to approximately 125×. The paper says its production GPU-hour denominator has not been verified; 106 of 7,062 Astra attempts also lacked reported usage, so the API total is a known-usage subtotal. The claimed savings concern video captioning, not Flora’s 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 Costs, Persistent Data Bottlenecks
Robot-learning systems need large collections of video annotated with what happened and when. Those labels help developers select examples, identify failed actions and train models to improve behavior. If automated labeling is substantially cheaper, it could reduce one expense in building training datasets and make it easier to process more footage.
That does not remove other constraints. The reported price is not a verified production bill, and companies cannot treat the release as a free commercial service under its stated license. Robot hardware, safety testing and access to representative data remain separate costs. Flora may offer researchers a useful tool and companies a point of comparison, but the published figure alone does not establish what a production deployment would cost.
As an affiliate, we earn on qualifying purchases.
Three Models for Robot Video
Flora targets three kinds of annotation described in the paper: timestamped captions for actions, step-level grades such as normal, suboptimal or failure, and scores for an overall attempt or a user-defined rubric. The segmenter finds where an action begins and ends; the captioner uses video context to write an instruction for each interval; the rewarder scores episodes and can grade actions at half-second intervals.
Nvidia tested the models on an H100 and says the captioner can accept a main camera feed, with optional wrist cameras. The paper’s benchmark results are mixed rather than a clear win across every task. Flora scored 57.1% on the paper’s overall comparison, ahead of GPT-6 Astra at 53.7%, Gemini 3.8 Flash at 38.3% and Claude Opus 5 at 30.4%. Nvidia reports Flora performed better on human videos and partial-success cases, while Astra remained stronger on robot videos, caption accuracy and rubric scoring.
On public WGO-Bench segmentation, Flora scored 50.3 against Astra’s 46.0. On Nvidia’s human-video subset, the figures were 37.4 and 29.7. On RoboRewardBench, the paper reports Flora’s error was 27.8% lower than Astra’s. These results come from different benchmarks and tasks, so the scores should not be read as a single universal measure of labeling quality.
“Research or evaluation purposes only.”
— Nvidia, describing the models’ license
Cost and Commercial Use Limits
The actual cost of running Flora at production scale is not established by the published comparison. Nvidia’s estimate assumes a $10 GPU-hour rate and modeled throughput, and excludes storage and overhead. The comparison also sets that estimate beside Astra API charges measured on a limited, vendor-selected set of episodes; the paper notes both the unverified GPU-hour denominator and missing API-usage reports.
Benchmark results do not settle how Flora will perform across different robots, camera setups, environments or labeling needs. Nvidia ran most of the tests described in the source material, and the quality measures differ by task. The license cards specify noncommercial research or evaluation use, with derivatives subject to the same restriction; the practical scope of permitted use should be checked against the license text. Nvidia may use the models commercially, according to the source material.
Independent Tests and License Review
The next useful evidence will be independent benchmark results and real-world operating costs from teams running Flora on their own video collections and hardware. Those tests could show whether the modeled GPU budget holds at different scales, and whether its relative strengths on human footage and partial successes translate to other robotics datasets.
Researchers and companies will also need to review the license before adopting the released weights or building derivatives. The release makes Flora available for research and evaluation, but the available information does not indicate a commercial license or a verified production-cost study. Nvidia’s benchmark and cost claims remain the main published basis for judging the system for now.
Key Questions
What is Nvidia Flora?
Flora is a group of three models for analyzing robot and human video: one identifies action intervals, one captions them, and one scores robot performance.
Is Flora really 125 times cheaper than GPT-6 Astra?
Nvidia reports a ratio of about 122.5 to one, rounded to approximately 125×. The comparison uses a modeled Flora GPU cost of $0.72 per video hour and measured Astra API charges of $88.49 per hour on a 50-episode test set, so it is not a direct comparison of two measured production bills.
Can businesses use Flora commercially?
The models are released under the NVIDIA OneWay Noncommercial License, which limits use to research or evaluation and passes that restriction to derivatives, according to the source material. Businesses should review the license terms; the release is not presented as a free commercial pipeline.
Does Flora outperform other models on every task?
No. Nvidia’s reported tests show strengths on human videos, action boundaries and partial-success cases, while GPT-6 Astra scores better on robot videos, caption accuracy and rubric scoring. Performance varies by benchmark and task.
What is not yet known about Flora?
Its production operating cost has not been verified, and the published results do not establish performance across all robots or real-world settings. Independent tests and more detailed deployment evidence are still needed.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
