🔍 Read the full analysis: Nvidia Flora’s Cost Advantage For Robot Video Labeling Comes With A Catch 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 released Flora, three models for segmenting, captioning and scoring robot-video tasks. 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, results vary by benchmark, and its 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 robot tasks, and estimates its captioning cost at $0.72 per video hour. That is far below the $88.49 per hour Nvidia reports for GPT-6 Astra on its test episodes, but the comparison pairs a modeled GPU cost with measured API usage, and Flora’s license allows research and evaluation—not commercial deployment.
Flora comprises Flora-TempSeg, a 9-billion-parameter model that identifies action intervals; Flora-Captioner, a 27-billion-parameter model that writes an instruction for each interval; and Flora-Rewarder, a 27-billion-parameter model that scores an episode or grades actions at half-second intervals. Nvidia says the models were used as a data engine for its GR00T-2 robot models. The project paper and model cards say Flora was trained on fewer than 10,000 hours of human and robot video.
Nvidia’s cost comparison uses 50 episodes from its DenseCaptionBench dataset, totaling 4,483 seconds of video, or about 75 minutes. The company reports GPT-6 Astra usage at $88.49 per video hour using public API rates dated October 4, 2026. Flora’s $0.72 estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput. Nvidia gives a range of $0.51 to $1.26 per video hour and says storage and overhead are excluded. The paper notes that the production GPU-hour denominator has not been verified.
The headline cost ratio is about 122.5 to one, rounded by Nvidia to approximately 125 times. The comparison covers captioning, not Flora’s reward model, and the Astra total is a known-usage subtotal: 106 of 7,062 attempts reported no usage. On the same episodes, Nvidia reports Flora’s score at 57.1% and Astra’s at 53.7%. These results come from tests Nvidia conducted, rather than an independent evaluation.
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, locate action boundaries and judge whether a robot succeeded. Automating that work could make it less expensive to prepare training data and review robot performance, especially when datasets contain many hours of recordings. Nvidia’s figures suggest a potentially large cost difference for captioning, even allowing for its more expensive modeled scenario.
But the estimate does not establish the cost of a full production pipeline. It uses a stated GPU rental assumption and modeled throughput, while excluding storage and overhead. Nor does the comparison show that Flora is better across the board: the reported overall score is close to Astra’s, and the benchmark results vary by task. For robotics developers, the release is both a possible research tool and a reference point for testing alternatives, rather than evidence that labeling costs have been settled at 72 cents per hour.
The license further limits the immediate business impact. Nvidia’s OneWay Noncommercial License permits research or evaluation purposes only; the source says derivatives inherit the restriction. That makes the models available for study, but does not give companies permission to use them as a commercial labeling service or production pipeline. Nvidia’s own commercial use does not change the stated terms for other users.
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 Handle Video Labels
Robot-learning systems use video of tasks such as moving objects, handling tools or folding fabric. Training data can require labels that mark when actions begin and end, describe each action in language, and assess whether a step or complete attempt went well. Flora separates those jobs across its three models instead of treating all annotation as one task.
According to Nvidia’s paper, TempSeg returns time intervals for individual operations. Captioner describes those intervals using surrounding video for context and can take a main camera feed, or a main feed plus wrist-camera views. Rewarder can give an episode a score from one to five, classify short segments as normal, suboptimal or failure, or score performance against a user-provided rubric. The models are based on Alibaba’s Qwen3.5 models, according to the project materials.
The release is not a lightweight local tool. Nvidia’s documentation describes testing on an H100 system running Linux. The source estimates that each 27-billion-parameter model needs about 55 GB for its weights in 16-bit precision, putting it in data-center hardware territory. That is a separate practical cost from the per-video-hour estimate and may affect who can run the models.
“expensive at the scale of modern robot datasets”
— Nvidia’s paper, describing frontier models
Cost and Performance Still Need Testing
The $0.72 figure is not a bill from a completed production run. It depends on assumed GPU rental prices and modeled throughput, and Nvidia says the production GPU-hour denominator has not been verified. The source also reports that 106 of Astra’s 7,062 attempts had no usage data, so its $88.49 figure is a known-usage subtotal. The small test set—50 episodes selected from Nvidia’s benchmark—does not show how costs or results would scale across other datasets, video lengths or operating conditions.
Performance is mixed across tasks. Nvidia reports Flora ahead on some measures, including human-video captioning and action segmentation, while Astra performs better on robot-video captioning, caption accuracy and rubric scoring in the paper’s comparisons. The supplied source does not provide enough detail to independently assess how the evaluations were conducted or how well results generalize beyond those tests. It is also unclear what costs would apply to a company seeking commercial rights or building a licensed production system.
Independent Tests and License Terms
The next useful evidence will be independent evaluations on broader robot-video datasets, with costs measured during actual runs and reported alongside hardware, throughput, storage and operational overhead. Further testing could clarify whether Flora’s reported advantages hold across different robots, camera setups and task types, and whether its cost advantage remains when the full annotation workflow is counted.
Organizations considering the models will also need to review the license before use. Under the terms described in the source, commercial users cannot treat the public release as a free production service. Nvidia may publish further performance or deployment details, but the source material does not specify a date for additional results or a change to the license.
Key Questions
What is Nvidia Flora?
Flora is a set of three models for robot-video labeling: one segments action intervals, one captions them, and one scores actions or complete episodes.
Is Flora proven to cost 72 cents per video hour?
No. Nvidia’s $0.72 estimate is modeled using an assumed GPU rental price and modeled throughput. The paper says the production GPU-hour denominator has not been verified; the estimate excludes storage and overhead.
How was the comparison with GPT-6 Astra made?
Nvidia compared Flora with Astra on 50 episodes totaling about 75 minutes. Astra’s $88.49-per-hour figure uses reported API usage and public rates dated October 4, 2026; Flora’s figure is a modeled GPU budget. The comparison covers captioning, not the Flora reward model.
Can companies use Flora commercially?
The source says the models use the NVIDIA OneWay Noncommercial License, which permits research or evaluation only, with derivatives subject to the same restriction. Companies should review the license terms before any use.
Does Flora outperform GPT-6 Astra on every test?
No. Nvidia reports Flora ahead on some measures, including parts of the human-video and action-segmentation evaluations, while Astra leads on several robot-video captioning and rubric-scoring measures. Results depend on the task and benchmark.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
