🔍 Read the full analysis: What Nvidia’s Flora Means For Robot Video Labeling Costs And Licensing on ThorstenMeyerAI.com
Get smart everyday buys 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, a set of three models 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 on its test set, but the Flora figure is modeled rather than a measured production bill. The models carry a noncommercial license, and benchmark results are mixed.
Nvidia’s GEAR robotics lab has released Flora, three models designed to divide robot videos into actions, describe those actions and score robot performance. Nvidia estimates that Flora’s captioning costs $0.72 per hour of video, versus $88.49 for GPT-6 Astra on the lab’s test set, but the Flora figure is a modeled GPU cost, not a measured production bill, and the models are licensed for research or evaluation only.
The Flora suite includes a 9-billion-parameter segmenter that identifies action intervals, a 27-billion-parameter captioner that writes an instruction for each interval, and a 27-billion-parameter reward model that scores an entire attempt or grades short steps. Nvidia says the models support robot-training data work, including the labeling pipeline for its GR00T-2 robot models. The paper says Flora was trained on fewer than 10,000 hours of human and robot video.
The cost comparison uses 50 episodes totaling 4,483 seconds—about 75 minutes—from Nvidia’s DenseCaptionBench. GPT-6 Astra’s $88.49 per video hour is based on reported API usage at public rates dated Oct. 4, 2026. Flora’s $0.72 figure is calculated from modeled throughput on eight GB300 GPUs, priced at an assumed $10 per GPU-hour. Nvidia gives a modeled range of $0.51 to $1.26 per video hour, excluding storage and other overhead, and says the production GPU-hour denominator has not been verified.
The paper reports an overall score of 57.1% for Flora on the tested episodes, compared with 53.7% for Astra, 38.3% for Gemini 3.8 Flash and 30.4% for Claude Opus 5. Results vary by task: Flora did better on human videos, action boundaries and partial-success cases, while Astra did better on robot-video captions and rubric scoring. Nvidia conducted most of the tests, and the cost ratio applies to captioning—not to Flora’s reward model.
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 identify what happened in a recording and whether a robot’s action succeeded. Lower labeling costs could make it less expensive to process large training sets, reducing one obstacle to developing robot-learning systems. Nvidia says Flora can process up to 200,000 hours of video per day using 512 GB300 GPUs; that is a company-reported capacity figure, not an independently verified deployment result.
The license changes the practical value for businesses. The models are released under Nvidia’s OneWay Noncommercial License, which permits research or evaluation, not commercial use; derivatives inherit the restriction. That means companies cannot treat the published models as a free production labeling service under the stated terms. They may still use Flora as a research benchmark or technical reference, but commercial deployment would require suitable rights or a different arrangement.
Nor does cheaper labeling solve the full cost of robotics. Teams still need video and robot hardware, data collection, compute, and methods to test safety and reliability. The reported price comparison is therefore a possible reduction in one data-processing expense, not a measure of total robot-training costs or a guarantee of better deployed performance.
Products Worth Considering
Apple iPad Air (4th Generation, 10.9-inch, Wi-Fi, 256GB) – Space Gray
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
Dell UltraSharp U2725QE 27-Inch 4K UHD IPS Black Monitor
As an affiliate, we earn on qualifying purchases.
Three Models for Video Labels
Robot-learning systems use labeled recordings to connect actions with outcomes. In the Flora paper’s description, timestamped captions identify actions and when they occur; process rewards grade short steps; and outcome rewards score an entire attempt. These labels can help teams organize training examples and identify failures. Flora packages all three functions into separate models rather than one general video tool.
The models are based on Alibaba’s Qwen3.5 family: Flora-TempSeg adapts Qwen3.5-9B, while the captioner and rewarder use 27-billion-parameter models. Nvidia’s model cards specify inference code in the Hugging Face repositories and list an Nvidia H100 on Linux as test hardware. The source material estimates that each 27-billion-parameter model requires roughly 55 GB for weights in 16-bit precision, placing local use closer to data-center hardware than a typical consumer PC.
Nvidia’s headline comparison is about captioning costs on its selected benchmark. Its paper reports a modeled cost ratio of about 122.5 to 1, rounded to approximately 125 times. That calculation divides Astra’s reported API cost by Flora’s modeled GPU cost; it is not a comparison of two measured production invoices.
Cost and License Boundaries
The reported $0.72 per hour is a modeled estimate based on an assumed $10 GPU-hour rate, not an observed bill from a production labeling operation. Nvidia lists a range and excludes storage and overhead; actual cost will depend on hardware pricing, throughput, utilization and deployment needs. The Astra figure also has a caveat: the paper says 106 of 7,062 attempts reported no usage, making its total a known-usage subtotal.
The benchmark covers 50 episodes selected for Nvidia’s DenseCaptionBench, and the supplied material does not establish how well those results generalize to other datasets or production workloads. The overall score does not erase the task-level differences: Flora leads on some categories while Astra leads on others. The cost ratio also excludes the reward model, so it cannot be read as the cost of all three Flora functions.
The noncommercial license is stated in the model materials, but the source does not describe a commercial licensing option or say whether Nvidia will offer one. It also does not provide independent verification of the claimed large-scale daily throughput. Those details matter to organizations weighing whether the models could be used beyond research and evaluation.
Testing Beyond Nvidia’s Benchmark
The next useful test is whether independent teams can reproduce the reported quality and throughput on different robot-video datasets and with clearly measured hardware costs. Such testing would help establish whether the modeled savings hold once deployment overhead, storage, data preparation and the full labeling workflow are included.
Organizations considering Flora should also review the OneWay Noncommercial License and model-card terms before building workflows around the release. Unless Nvidia publishes different terms or a separate commercial agreement, the stated permission remains limited to research and evaluation. Nvidia has not, in the supplied material, announced such an alternative.
For now, Flora is evidence that Nvidia is offering a lower-cost research approach to robot-video labeling, not proof that commercial labeling costs have fallen to 72 cents per hour. The benchmark’s limits, the model-by-model quality differences and licensing terms will shape how far the release can be used in practice.
Key Questions
What is Nvidia Flora?
Flora is a set of three video models from Nvidia’s GEAR robotics lab: one segments actions by time, one captions those actions, and one scores robot attempts or steps.
Is Flora’s $0.72 cost a measured production price?
No. Nvidia’s $0.72 per video hour is modeled using assumed GPU pricing and throughput. The paper gives a range of $0.51 to $1.26 and says the production GPU-hour denominator has not been verified.
Can a company use Flora commercially?
The stated OneWay Noncommercial License allows research or evaluation only, and the source says derivatives inherit that restriction. The supplied material does not identify a commercial-use option.
Did Flora outperform GPT-6 Astra in every test?
No. Flora scored higher overall on Nvidia’s reported test set and did better on human videos, action boundaries and partial-success cases. Astra performed better on robot-video captions and rubric scoring.
Does the 125-times figure cover every Flora model?
No. The comparison concerns captioning costs on a 50-episode benchmark. The reward model is not included in the reported ratio, and the comparison pairs a measured API-usage subtotal with a modeled GPU cost.
Source: ThorstenMeyerAI.com
