🔍 Read the full analysis: Nvidia’s Flora Promises Cheaper Robot Video Labeling; Licensing Shapes The Deal on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Nvidia’s GEAR robotics lab released Flora, three models for segmenting, captioning and scoring robot videos. Nvidia estimates labeling costs at $0.72 per video hour, compared with $88.49 for GPT-6 Astra on a small test set, but Flora’s figure is modeled 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 robot video, and estimates that one part of the work could cost $0.72 per video hour. That estimate is far below the $88.49 Nvidia reports for GPT-6 Astra on the same test episodes, but the comparison pairs a modeled GPU cost with measured API usage, and Flora’s license permits research and evaluation—not commercial deployment.
Flora is made up of Flora-TempSeg, a 9-billion-parameter model that identifies action intervals; Flora-Captioner, a 27-billion-parameter model that writes instructions for those intervals; and Flora-Rewarder, also 27 billion parameters, which scores whole attempts or individual steps. The models are based on Alibaba’s Qwen3.5 family and are available through Hugging Face, according to the project materials.
Nvidia’s cost comparison uses 50 episodes totaling 4,483 seconds, or about 75 minutes, from its DenseCaptionBench. GPT-6 Astra’s $88.49 per video hour is based on reported API usage and public rates cited in the paper. Flora’s $0.72 is not a measured invoice: it is calculated from modeled throughput on eight GB300 GPUs priced at an assumed $10 per GPU-hour. Nvidia lists scenarios ranging from $0.51 to $1.26 per hour, excluding storage and other overhead.
The paper reports an overall benchmark score of 57.1% for Flora and 53.7% for Astra on those episodes. Results vary by task: Flora performed better on human videos, action boundaries and some partial-failure judgments, while Astra scored higher on robot videos, caption accuracy and rubric scoring. Nvidia says it uses Flora as a data engine for its GR00T-2 robot models and 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 measurement.
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 robot developers identify what an action is, when it happens and whether it succeeds. Automating that work could reduce one expense in preparing data for robot foundation models, which need examples of tasks such as placing objects or handling tools. If the reported cost advantage holds at larger scale, teams could label more recordings for a given budget.
But the cost comparison is not an apples-to-apples production test. It compares a modeled GPU budget with API charges on a limited, Nvidia-selected benchmark, and the 125-times figure concerns captioning rather than every Flora function. Quality is also mixed rather than uniformly higher. The release therefore offers evidence of a potentially cheaper approach, not proof that it will replace commercial labeling services across real-world datasets.
Access is another constraint. The models use the NVIDIA OneWay Noncommercial License, which allows research or evaluation purposes, while derivatives inherit the restriction. A company cannot treat the released weights as a free commercial labeling pipeline under those terms. For commercial users, Flora may serve as a research reference or a technical starting point, but the license limits direct business use.
Popular right now

De'Longhi La Specialista Arte Evo Semi-Automatic Espresso Machine with Grinder and Cold Brew

Breville Barista Express Espresso Machine with Grinder, Black Truffle

Dyson Cyclone V10 Origin Cordless Vacuum
As an affiliate, we earn on qualifying purchases.
Three Models for Robot Video
Robot-learning datasets include recordings from robot cameras and human demonstrations. The Flora paper describes three related labeling needs: timestamped captions that identify actions, step-level rewards that mark normal, suboptimal or failed behavior, and scores for an entire attempt. Flora splits those tasks among its three models rather than relying on one general-purpose model.
The model cards say Flora was trained on fewer than 10,000 hours of human and robot video and used in preparing data for GR00T-2. The paper’s comparison includes other language models as well as Astra, but its headline cost ratio is based on the Astra comparison. Nvidia’s appendix says the GPU-hour denominator used for its production cost estimate has not been verified. Running the two 27-billion-parameter models also calls for substantial data-center memory, rather than typical consumer hardware.
Cost and Quality Still Need Testing
The cost estimate has not been confirmed in a production deployment, and the source material does not provide an independently audited GPU throughput figure. It is also unclear how the results would change across larger or more varied datasets, different video formats, and the overhead of storing and managing data. The benchmark covers 50 episodes and was assembled by Nvidia.
Performance depends on the task: Flora’s stronger results on human footage and partial success do not establish that it is better for all robot recordings. The comparison also excludes the cost of using Flora-Rewarder from the stated 125-times captioning claim. Finally, Nvidia’s claimed processing capacity and the practical results at that scale have not been independently verified. The noncommercial license’s limits mean commercial availability cannot be inferred from the models’ public download access.
Production Tests and License Terms
The next useful evidence would be independent tests on larger and more varied robot datasets, with measured GPU costs that include operational overhead. Those tests could show whether Flora’s modeled savings persist outside Nvidia’s benchmark and whether its task-specific strengths translate to production labeling.
Organizations considering the models will also need to review the license before using them beyond research or evaluation. Nvidia has not, in the supplied materials, announced commercial licensing terms or a verified production-cost figure. Until those details emerge, the reported cost advantage remains a modeled estimate, and commercial users must account for the stated restriction.
Key Questions
What is Nvidia Flora?
Flora is a set of three video models for identifying action intervals, writing captions and scoring robot attempts or individual steps.
Is the $0.72-per-hour figure a measured cost?
No. Nvidia presents it as a modeled estimate based on assumed GPU pricing and throughput. The paper says the production GPU-hour denominator has not been verified.
What does Nvidia mean by about 125 times cheaper?
The ratio comes from dividing the reported $88.49 Astra API cost by Flora’s modeled $0.72 cost per video hour. It is based on 50 benchmark episodes and applies to the captioning comparison, not every Flora model or cost in a production system.
Can companies use Flora commercially?
The models are released under the NVIDIA OneWay Noncommercial License, which permits research or evaluation. The supplied material says derivatives inherit that restriction; it does not describe commercial-use terms.
Did Flora outperform GPT-6 Astra overall?
On Nvidia’s reported benchmark, Flora scored 57.1% versus Astra’s 53.7%. The results were mixed by task: Astra led on robot videos, caption accuracy and rubric scoring, while Flora led on some human-video, segmentation and partial-success measures.
Source: ThorstenMeyerAI.com
