🔍 Read the full analysis: How Flora Changes The Economics Of Robot Video Labeling—and Its License Terms 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, a three-model system for segmenting, captioning and scoring robot videos. Nvidia estimates captioning costs at $0.72 per video hour, compared with $88.49 for GPT-6 Astra, but Flora’s figure is modeled rather than a measured production bill, and its license restricts use to research or evaluation. Benchmark results are mixed, with Flora ahead in some tasks and Astra ahead in others.
Nvidia’s GEAR robotics lab has released Flora, a set of three models designed to segment, caption and score robot videos, and estimates that its captioning can cost $0.72 per video hour. The estimate is not a measured production bill, and the models are released under a noncommercial license, limiting their direct use by companies seeking a production labeling pipeline.
Flora includes a 9-billion-parameter segmenter that identifies action intervals, a 27-billion-parameter captioner that describes actions in English, and a 27-billion-parameter reward model that scores overall episodes or individual steps. According to the model cards, the system was used as a data engine for Nvidia’s GR00T-2 robot models and was trained on fewer than 10,000 hours of human and robot video. The models are built on Alibaba’s Qwen3.5 family.
Nvidia’s paper compares Flora’s estimated captioning cost with usage of GPT-6 Astra on the same 50 test episodes, totaling 4,483 seconds, or about 75 minutes, of video. Astra’s reported API usage cost was $88.49 per video hour. Flora’s $0.72 figure is modeled using eight GB300 GPUs at an assumed rate of $10 per GPU-hour and modeled throughput; Nvidia gives a range of $0.51 to $1.26, excluding storage and overhead. The paper says the production GPU-hour denominator has not been verified.
The resulting ratio is about 122.5 to one, rounded by Nvidia to approximately 125 times cheaper. That comparison concerns captioning, not all three Flora models. The reported Astra total also excludes 106 of 7,062 attempts for which no usage was reported, making it a known-usage subtotal. Nvidia’s benchmark results on the test set give Flora a score of 57.1%, versus 53.7% for Astra, but the paper’s wider results show a mixed performance picture rather than a consistent Flora lead.
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 happened during a task, where each action began and ended, and whether a step succeeded. They can support dataset filtering, training and evaluation. If Flora’s projected costs hold at scale, lower labeling expenses could ease one part of the data workload for robotics teams, where generating useful training material can be time-consuming and expensive.
But the economic comparison does not show what a company would actually pay to deploy Flora. Its estimate depends on GPU rental pricing, throughput and operational costs, and it excludes storage and overhead. The license is another practical limit: the NVIDIA OneWay Noncommercial License allows research or evaluation use, not ordinary commercial production. For businesses, Flora may offer a reference point for capabilities and costs, but the cited terms do not provide a free commercial labeling service.
Cheaper annotation would not resolve other constraints on robotics, including collecting representative robot data, hardware expense and safety. Flora addresses a specific part of the development process: turning video into labels and scores. Its value to users and developers will depend on both the quality of those outputs and whether future terms allow commercial deployment.
Popular right now
TCL 85 Designer Series QD-Mini LED TV (Model 85A400U) with Google Smart TV
Shark Navigator Robot Vacuum with Self-Empty Base, Dark Grey (RV2120AE)
iRobot Roomba 105 Vac Robot Vacuum
As an affiliate, we earn on qualifying purchases.
Three Models for Robot Video
Robot learning systems use video of tasks such as placing objects, handling wires or folding fabric. As described in the Flora paper, training and evaluation can use timestamped captions, which tie instructions to moments in a video, as well as rewards that grade steps or an entire attempt. The segmenter locates action boundaries; the captioner describes each interval; and the rewarder can score an episode from one to five, label half-second steps as normal, suboptimal or failure, or apply a user-provided rubric.
The models are available through their Hugging Face repositories, with inference code included there. Nvidia’s documentation says testing used an H100 GPU on Linux. The two 27-billion-parameter models require substantial memory for 16-bit weights, according to the source material, so running them locally is a data-center task rather than a typical consumer-computer workload.
The cost comparison uses a fixed benchmark of 50 episodes selected from Nvidia’s DenseCaptionBench: 20 with human hands, 15 with two-finger grippers and 15 with dexterous robot hands. Nvidia reports Flora leading on some measures, including action-boundary segmentation on WGO-Bench and the human-video subset of its own benchmark. Astra leads on robot-video caption accuracy and rubric scoring. On the public RoboRewardBench, Flora’s reported error is 27.8% lower than Astra’s, while the overall results vary by task.
“Research or evaluation purposes only.”
— NVIDIA OneWay Noncommercial License
Costs and Commercial Rights
The $0.72 figure is a modeled estimate, not an invoice from a production deployment. Nvidia’s paper lists a scenario range and says the production GPU-hour denominator has not been verified. Actual costs could depend on throughput, GPU availability, storage, engineering and other overhead not included in the calculation. The comparison is also based on a small, fixed set of 50 episodes selected by Nvidia, and the reported figures compare modeled GPU costs with measured API usage.
It is not clear from the supplied material whether Nvidia plans to offer a commercial license, what terms such a license might carry, or whether future Flora versions will use different terms. Nvidia itself may use the models commercially under its own rights, but that does not grant the same permission to outside users. Benchmark scores also do not establish performance across other robots, video conditions or real-world production datasets.
License and Deployment Questions
The immediate next step for researchers is to examine the model cards, paper and license, then test Flora on their own videos and hardware. Those evaluations can show whether the reported strengths transfer to data beyond Nvidia’s benchmark and what operating costs emerge under local conditions. Companies considering deployment will need to resolve the noncommercial-use restriction before treating Flora as a production option.
Further clarity would come from verified throughput and cost measurements in a production setting, expanded independent evaluations, and any licensing update from Nvidia. Until then, the headline cost ratio is best understood as a comparison of one modeled Flora captioning scenario with known GPT-6 Astra API usage on a limited test set—not as a general price guarantee for robot-video labeling.
Key Questions
What does Nvidia Flora do?
Flora is a set of three models for robot video: a segmenter that marks action intervals, a captioner that describes them, and a rewarder that scores steps or full episodes.
Is Flora really 125 times cheaper than GPT-6 Astra?
Nvidia’s paper estimates a ratio of about 122.5 to one, rounded to approximately 125 times, for captioning on 50 test episodes. Flora’s $0.72-per-video-hour cost is modeled, while Astra’s $88.49 figure comes from reported API usage; the comparison is not a verified production-cost audit.
Can companies use Flora commercially?
The supplied license information says the models are for research or evaluation purposes only. Commercial production use is not permitted under those terms, and the material does not say whether Nvidia will offer a separate commercial license.
Does Flora outperform Astra on every benchmark?
No. The supplied results show Flora ahead on some measures, including human-video labeling and action boundaries, while Astra performs better on robot-video caption accuracy and rubric scoring. Results depend on the task and benchmark.
Source: ThorstenMeyerAI.com
