🔍 Read the full analysis: Robot Video Labeling And Nvidia Flora: How Much Does The License Change? on ThorstenMeyerAI.com
Get hardware and tech essentials 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 released Flora, a set of three models for segmenting, describing and scoring robot videos. Nvidia estimates labeling costs at $0.72 per video hour versus $88.49 for GPT-6 Astra on its test, but Flora’s figure is modeled, its benchmarks show mixed results, and its license permits research or evaluation use rather than commercial deployment.
Nvidia’s GEAR robotics lab has released Flora, three models designed to label robot-training videos, while restricting use to research or evaluation purposes under a noncommercial license. Nvidia estimates Flora’s captioning cost at $0.72 per hour of video, compared with $88.49 for GPT-6 Astra in a test it describes in a research paper; the Flora figure is a model, not a measured production bill.
Flora comprises a 9-billion-parameter segmenter that identifies the start and end of actions, a 27-billion-parameter captioner that describes each interval, and a 27-billion-parameter reward model that grades robot behavior. The models are based on Alibaba’s Qwen3.5 family. Nvidia says Flora was used as a data engine for its GR00T-2 robot model and was trained on fewer than 10,000 hours of human and robot video.
The comparison behind the cost claim used 50 episodes totaling 4,483 seconds, or about 75 minutes, from Nvidia’s DenseCaptionBench. Nvidia reports GPT-6 Astra’s API usage cost as $88.49 per video hour. Flora’s $0.72 figure is calculated from an assumed rate of $10 per GPU-hour and modeled throughput using eight GB300 GPUs. Nvidia gives a range of $0.51 to $1.26 per video hour and says the production GPU-hour denominator has not been verified. The ratio of the two headline figures is about 122.5, rounded by Nvidia to approximately 125 times.
On that same test, Flora scored 57.1% and GPT-6 Astra 53.7%, according to the source report’s account of the paper. The reported comparison excludes the reward model and concerns captioning. Other results are not uniformly in Flora’s favor: the paper reports advantages on human video and partial-success cases, while Astra performs better on robot-video caption accuracy and rubric scoring. Nvidia conducted most of the tests, and the results do not establish performance across all robot datasets or operating conditions.
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 Commercial Use
Video annotation is one step in preparing the data used to train and evaluate robot models. If Flora can deliver useful labels at lower computing cost, robotics researchers may be able to process more video or reduce reliance on paid third-party APIs. That could ease one part of building training datasets, but it does not resolve the separate costs and challenges of collecting robot footage, operating hardware, or testing safety.
The license limits the immediate business case. The three models are distributed under the NVIDIA OneWay Noncommercial License, which allows research or evaluation use; the source says derivatives carry the same restriction. Companies cannot treat the release as an unrestricted, free production pipeline under those terms. Nvidia’s own commercial use does not grant other organizations the same rights. The cost estimate is also not a like-for-like measured comparison: it contrasts a modeled GPU budget with recorded API usage.
For developers, the release can still offer a research tool and a point of comparison for video-labeling quality. But teams considering commercial deployment would need to resolve licensing rights and validate costs and performance on their own workloads. The claimed savings should not be read as proof that end-to-end robot development has become cheaper by the same ratio.
NVIDIA GPU workstation for AI video processing
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Three Models for Video Labels
Robot-learning datasets contain recordings of tasks such as placing a mug, passing an object between hands or folding fabric. To make such footage useful for training, teams may need timestamped descriptions of actions, scores for each step, and an assessment of the overall attempt. Those labels can help filter examples, identify mistakes and monitor outcomes.
Flora assigns these jobs across three models. TempSeg proposes action intervals; the captioner writes an instruction for each interval and can use multiple camera views; the rewarder can score a full episode from 1 to 5, classify behavior at half-second intervals, or apply a user-provided rubric. The model cards and paper are available through Hugging Face, and the source report says the inference code is included in those repositories. Its described testing used an Nvidia H100 on Linux; the report estimates that each 27-billion-parameter model requires about 55 GB for weights in 16-bit precision, suggesting substantial hardware needs.
The paper’s comparison also included Gemini 3.8 Flash at $8.48 per video hour and Claude Opus 5 at $41.70, with reported scores of 38.3% and 30.4%, respectively. Those figures are tied to the same selected test set and reported methodology; they do not establish universal rankings. The estimated Flora cost excludes storage and other overhead, while the Astra API total included 106 of 7,062 attempts for which usage was not reported, making it a known-usage subtotal.
NVIDIA Blackwell GPU server for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Cost and Benchmark Limits
The $0.72-per-hour result is a modeled estimate, not a production invoice. Nvidia’s calculation assumes GPU pricing and throughput, and the paper says its production GPU-hour denominator has not been verified. The comparison also uses a short, fixed test set selected from Nvidia’s benchmark, so it remains unclear how the cost and quality figures would change across larger or independently selected datasets.
Quality results vary by task and video type, and most of the testing was conducted by Nvidia, according to the source report. The source does not provide independent replication of the headline results. It is also unclear whether Nvidia will offer a separate commercial license, what terms might apply, or whether users can obtain rights beyond research and evaluation. The stated 125-times comparison applies to captioning costs, not all three Flora models or the full cost of preparing robot-training data.
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 Terms
The immediate next step for researchers is to test the released models against their own videos, hardware budgets and evaluation criteria, while checking the license before using outputs or derivatives in commercial work. Further reporting or documentation could clarify how throughput was modeled, how the cost behaves at production scale, and how well results transfer beyond the 50 episodes in Nvidia’s comparison.
For companies, any commercial use depends on obtaining appropriate rights rather than assuming the research release permits deployment. Nvidia has not, in the source material, announced a commercial license or independently verified the production cost calculation. Until those questions are addressed, Flora’s figures are best treated as reported research results and a potential benchmark, not a guaranteed operating cost.
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 video models for identifying action intervals, writing captions and scoring robot behavior. Nvidia says it was used to prepare data for its GR00T-2 model.
Does Flora really cost $0.72 per video hour?
That is Nvidia’s modeled estimate, based on an assumed $10 GPU-hour rate and modeled throughput using eight GB300 GPUs. It is not a measured production bill, and Nvidia lists scenarios from $0.51 to $1.26 per video hour.
What does the Flora license permit?
The source report says the NVIDIA OneWay Noncommercial License permits use for research or evaluation. It says derivatives inherit that limit, so the release should not be treated as permission for ordinary commercial deployment.
Did Flora outperform GPT-6 Astra in every test?
No. The reported results are mixed. Flora scored higher on Nvidia’s selected captioning test and had reported strengths on human videos and partial-success cases, while Astra did better on some robot-video captioning and rubric-scoring measures.
Does the 125-times figure cover all of Flora?
No. The ratio compares a modeled Flora captioning cost with GPT-6 Astra API usage on a 50-episode test. It does not include Flora’s reward model or all costs of collecting and preparing robot data.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
