🔍 Read the full analysis: Nvidia’s Flora Could Make Robot Video Labels 125× Cheaper, Subject To Its License on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- 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 about $0.72 per video hour versus $88.49 for GPT-6 Astra, but the comparison uses modeled GPU costs against measured API use, and Flora’s license permits research and evaluation, not commercial deployment.
Nvidia’s GEAR robotics lab has released Flora, a set of three models that segment, caption and score robot videos, and estimates that its captioning pipeline could cost about $0.72 per video hour—roughly 122.5 times below the measured GPT-6 Astra API cost in its test. The comparison is not like-for-like: Flora’s figure is a modeled GPU budget, and its noncommercial license limits use to research or evaluation.
Flora contains three models built from Alibaba’s Qwen3.5 family, according to the paper and model cards. Flora-TempSeg, with 9 billion parameters, 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 episode overall, rates steps as normal, suboptimal or failure, or evaluates performance against a user-provided rubric.
Nvidia’s cost comparison uses 50 episodes totaling 4,483 seconds, or about 75 minutes, from its DenseCaptionBench. GPT-6 Astra’s cost was calculated from reported API usage at public rates dated October 4, 2026: $88.49 per video hour. Nvidia estimates Flora at $0.72 per hour, using eight GB300 GPUs and an assumed price of $10 per GPU-hour. The paper gives a range of $0.51 to $1.26 and says storage and other overhead are excluded.
The resulting ratio is 122.5, rounded by Nvidia to approximately 125 times. The paper also says the production GPU-hour denominator has not been verified. The comparison covers captioning, not Flora’s reward model. Nvidia reports a 57.1% score for Flora on the test, compared with 53.7% for Astra, but the results reflect a vendor-selected benchmark and Nvidia conducted most of the evaluations itself.
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 Access
Video labels help robotics teams connect what a camera records with what a robot should do, identify when an action went wrong, and evaluate success. Reducing the expense of creating those labels could make it easier to process large training datasets. Nvidia says Flora is used as a data engine for its GR00T-2 robot models and that its system can process up to 200,000 hours of video per day on 512 GB300 GPUs. That capacity figure is Nvidia’s claim, not an independently verified production result.
For other organizations, the license is a major limit on the immediate commercial value. The three models are released under the NVIDIA OneWay Noncommercial License, which permits research or evaluation; derivatives inherit the restriction. Nvidia may use the models commercially, according to the source material. Companies considering a production labeling pipeline cannot treat Flora’s release as a free commercial tool. The cost estimate is best read as a possible efficiency benchmark, rather than a price businesses can currently rely on.
Cheaper labeling would address one part of the cost of building robot systems, not the entire problem. Teams still need useful video data, physical robots, safe testing and reliable performance outside a benchmark. The reported result also does not establish that lower-cost labels by themselves lead to better robots.
NVIDIA DGX workstation for AI model inference
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Flora Labels Robot Footage
Robotics teams use video of people and robots performing tasks to train and evaluate models. The Flora paper describes three types of labels: timestamped captions that identify an action and when it happens, step-level process rewards that mark normal or failed behavior, and outcome scores for an entire attempt. These can help researchers find relevant examples, examine errors and measure performance.
Flora’s models handle different parts of that work rather than providing a single label. The captioner can use one camera or a main camera alongside wrist cameras, while the rewarder can assess either progress through an episode or a rubric. The model cards say Flora was trained on fewer than 10,000 hours of robot and human video. The source reports that inference code is included in the Hugging Face repositories and that testing used an Nvidia H100 on Linux. It estimates that each 27-billion-parameter model needs about 55 GB for its weights at 16-bit precision, putting local use beyond typical gaming PCs.
Nvidia’s reported benchmark results are mixed. Flora scored higher than Astra on the overall DenseCaptionBench test and on its human-video subset, and the paper reports stronger results for action-boundary detection and partial-success assessment. Astra performed better on robot videos, caption accuracy and rubric scoring, according to the same source. These results do not show that one model is best for every video-labeling task.
“research or evaluation purposes only”
— NVIDIA OneWay Noncommercial License
NVIDIA H100 GPU server for robotics AI research
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Cost and Benchmark Caveats
The exact cost of running Flora in a production setting is not established by the paper. Nvidia’s $0.72 figure assumes a $10 hourly rate for each GB300 GPU and modeled throughput; it is not a measured invoice. The estimate excludes storage and overhead, and the paper explicitly says its production GPU-hour denominator is unverified. Astra’s comparison figure, by contrast, comes from reported API usage. The test also includes 106 of 7,062 attempts with no reported usage, making the $88.49 figure a known-usage subtotal, according to the source.
How well the results generalize beyond the 50 episodes is also unclear. The material describes the set as Nvidia’s own benchmark, and the evaluations were largely run by Nvidia. The supplied source does not give independent replication of either the cost estimate or benchmark scores. It also does not establish commercial rights for users; the stated license is noncommercial. Broader availability, licensing changes and independently measured operating costs remain unconfirmed.
robotics research workstation with multiple GPUs
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Commercial Use and Replication
The immediate next questions are whether independent researchers can reproduce Nvidia’s reported quality results and whether real deployments achieve similar throughput and cost. Organizations will also need to review the license before using the models or derivatives in products or commercial data pipelines. The source material does not announce a commercial license or a schedule for one.
Nvidia’s release gives robotics researchers a set of models to examine and compare with existing labeling tools. But the headline cost ratio should not be treated as a confirmed production saving: it rests on a modeled GPU budget, a limited benchmark and a noncommercial release. Further measurements, licensing information and results on a wider range of robot footage would clarify how far Flora can be used beyond research and evaluation.
robot camera system for machine learning video data collection
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 three-model system from Nvidia’s GEAR robotics lab for segmenting actions in video, writing captions for them and scoring robot performance.
Is Flora really 125 times cheaper than GPT-6 Astra?
That is Nvidia’s rounded estimate for captioning on its 50-episode test. Flora’s $0.72 per video hour is modeled, while Astra’s $88.49 figure is based on measured API usage, so the comparison is not like-for-like and is not a verified production price.
Can a company use Flora commercially?
The released models are under the NVIDIA OneWay Noncommercial License, which the source describes as allowing research or evaluation only. The material does not announce commercial-use terms.
Does Flora outperform GPT-6 Astra on every test?
No. Nvidia’s reported results favor Flora on some measures, including human videos and action boundaries, while Astra scores better on robot videos, caption accuracy and rubric scoring.
What costs are not included in Nvidia’s estimate?
The $0.72 estimate is based on assumed GPU pricing and modeled throughput. Nvidia’s paper says it excludes storage and overhead and that the production GPU-hour denominator has not been verified.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
