🔍 Read the full analysis: Nvidia’s Flora Could Change Robot Video Labeling Costs, But Check The License on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
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, compared with $88.49 for GPT-6 Astra on a small test set, but the Flora figure is modeled, quality results are mixed, and the models’ license restricts use to research or evaluation.
Nvidia’s GEAR robotics lab has released Flora, a set of three models that segment, caption and score video of people and robots performing tasks. Nvidia estimates one part of the system, video captioning, would cost $0.72 per hour of video to run, compared with $88.49 for GPT-6 Astra in its test; the estimate is modeled rather than a measured production bill, and Flora’s license limits use to research or evaluation.
The three models handle distinct labeling jobs. Flora-TempSeg, with 9 billion parameters, identifies the time boundaries of actions. The 27-billion-parameter Flora-Captioner writes an instruction for each segment, while the 27-billion-parameter Flora-Rewarder scores an episode overall, grades steps, or evaluates performance against a user’s rubric. Nvidia says the models were built on Alibaba’s Qwen3.5 models and used as a data engine for its GR00T-2 robotics work.
The headline cost comparison comes from Nvidia’s paper and its Appendix D. The benchmark used 50 episodes totaling 4,483 seconds, or roughly 75 minutes, from Nvidia’s DenseCaptionBench. GPT-6 Astra’s $88.49 per video hour is based on reported API usage at public rates dated October 4, 2026. Flora’s $0.72 figure is calculated using a modeled throughput on eight GB300 GPUs, with an assumed cost of $10 per GPU-hour. Nvidia gives a range of $0.51 to $1.26 per video hour, excluding storage and other overhead.
The paper reports a 57.1% score for Flora and 53.7% for Astra on the same test episodes. The comparison is not a complete measure of robotics performance: it covers captioning, not the cost of Flora’s reward model, and the test set was selected by Nvidia. The paper also says 106 of 7,062 Astra attempts reported no usage, making its cost total a known-usage subtotal. Nvidia’s stated cost ratio is about 122.5, rounded to approximately 125 times.
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 labels help robotics teams identify what happened during a task, when an action began and whether a step succeeded. Lower costs could let researchers process more recorded demonstrations or apply labels to data that would otherwise be too expensive to review with paid services. That may reduce one expense in training robot systems, but it does not by itself establish that robots will become cheaper or more capable.
The license is a direct constraint for companies weighing the models as infrastructure. All three are distributed under the NVIDIA OneWay Noncommercial License, which allows research or evaluation use; the source material says derivatives inherit the restriction. That makes Flora available for study and comparison, but not a straightforward free production pipeline for commercial labeling. The large cost gap also needs careful reading because it compares a modeled GPU budget with a measured API usage figure, rather than two observed production bills.
Products Worth Considering
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
DJI Osmo Mobile 8 Standard Combo Phone Gimbal Stabilizer
As an affiliate, we earn on qualifying purchases.
Three Models for Robot Video
Robotics datasets can include recordings from robot cameras, wrist cameras and human demonstrations. A labeling system may need to mark where each operation starts and ends, describe it in language, and assign scores to individual steps or a whole attempt. Those labels help researchers organize training examples and examine failures; they do not replace collecting the recordings or testing robots in physical settings.
According to the paper and model cards described in the source, Flora was trained on fewer than 10,000 hours of human and robot video. Its models and inference code are provided through Hugging Face repositories, and the reported test hardware for inference is an Nvidia H100 running Linux. The source estimates that each 27-billion-parameter model requires about 55 GB for weights in 16-bit precision, suggesting that local use requires data-center-class hardware rather than a typical consumer computer.
Nvidia’s reported benchmark results are mixed. Flora scored better on some measures, including action segmentation on WGO-Bench and the human-video subset of DenseCaptionBench, and the source reports a lower error than Astra on RoboRewardBench for partial-success judgments. Astra scored higher on some robot-video, caption-accuracy and rubric-scoring tests. These findings come mainly from evaluations Nvidia conducted, so they are evidence about the reported benchmarks, not a settled ranking across robotics tasks.
Cost and Benchmark Limits
The actual cost of running Flora at production scale is not established by the reported estimate. Nvidia’s calculation assumes a GPU-hour price and modeled throughput, and the paper says the production GPU-hour denominator has not been verified. The estimate also excludes storage and overhead. The test dataset is relatively small and comes from Nvidia’s own benchmark, so results may differ on other types of video, tasks or hardware.
It is also not clear from the supplied material how the license applies to every possible deployment or derivative beyond the stated noncommercial restriction. Readers and organizations should consult the license itself before use. The reported results do not establish that Flora outperforms Astra across all video-labeling tasks, and the 125-times comparison applies only to the captioning cost calculation, not the full cost of a robotics data pipeline.
Commercial Use and Broader Testing
The next useful evidence will be independent evaluations on broader robot-video datasets and observed costs from production deployments, including hardware utilization, storage and operational overhead. That would show whether the modeled savings hold beyond Nvidia’s selected test and assumptions.
For now, researchers can examine the model repositories and license, while companies need to check whether their intended use falls within the permitted research or evaluation scope. Further details about commercial permissions, deployment costs and performance outside the reported benchmarks remain the key points to watch.
Key Questions
What is Nvidia Flora?
Flora is a set of three video models for identifying action segments, writing captions for them and scoring robot performance. Nvidia’s GEAR lab released the models for robotics research and evaluation.
Is Flora really 125 times cheaper than GPT-6 Astra?
That is Nvidia’s rounded comparison for captioning on its test: $0.72 per video hour for Flora, modeled, against $88.49 for Astra based on reported API usage. The figures use different cost methods, and the result does not cover Flora’s reward model or full production overhead.
Can a company use Flora commercially?
The supplied source says the models use the NVIDIA OneWay Noncommercial License, restricted to research or evaluation, with derivative works subject to the same limit. Businesses should review the license and seek clarification before relying on Flora for commercial operations.
Does Flora outperform GPT-6 Astra?
Not across every reported test. Flora scored higher on some measures, including human-video labeling and certain action or reward benchmarks, while Astra did better on some robot-video, caption-accuracy and rubric-scoring evaluations. Most tests were run by Nvidia.
Source: ThorstenMeyerAI.com
