🔍 Read the full analysis: A 125× Price Shift For Robot Video Labeling? Nvidia Flora Has A License Catch on ThorstenMeyerAI.com
Get everyday 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 lab released Flora, three models for segmenting, captioning and scoring robot video. Nvidia estimates captioning costs 72 cents per video hour versus $88.49 for GPT-6 Astra, but the comparison mixes a modeled GPU budget with measured API usage, and Flora’s license is limited to research or evaluation.
Nvidia’s GEAR robotics lab has released Flora, a set of three models that segment, caption and score video for robot training, alongside an estimate that its captioning costs about $0.72 per video hour versus $88.49 using GPT-6 Astra. The comparison is not like-for-like: Nvidia’s Flora figure is a modeled GPU cost, while Astra’s is based on measured API usage, and Flora’s noncommercial license restricts use to research or evaluation.
Flora comprises Flora-TempSeg, a 9-billion-parameter model that identifies action intervals; Flora-Captioner, a 27-billion-parameter model that writes instructions for each interval; and Flora-Rewarder, a 27-billion-parameter model that grades an episode or individual steps. The project is from Nvidia’s GEAR lab and is built on Alibaba’s Qwen3.5 models. Nvidia says Flora was used as a data engine for its GR00T-2 robot models.
Nvidia’s cost comparison uses 50 episodes totaling 4,483 seconds, or about 75 minutes, from its DenseCaptionBench. The paper reports an Astra API cost of $88.49 per video hour and estimates Flora at 72 cents per hour, producing a ratio of about 122.5 to 1, rounded to approximately 125×. Flora’s estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput; Nvidia gives a range of $0.51 to $1.26 per hour and says storage and overhead are excluded.
On the same benchmark, Nvidia reports an overall score of 57.1% for Flora and 53.7% for Astra. Results vary by task: Flora scores better on human videos, action boundaries and partial-success cases, while Astra performs better on robot videos, caption accuracy and rubric scoring, according to the paper. The models are available through Hugging Face, with inference code in their repositories.
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
The Cost Claim Has Limits
Video labels help robotics teams turn recordings into training examples: captions describe actions, timestamps locate them, and reward scores identify successful or failed steps. Automating that work could lower the cost of preparing data for robot-learning systems. Nvidia says its system can process up to 200,000 hours of video per day on 512 GB300 GPUs, a company-reported capacity figure rather than an independent production measurement.
For businesses, the license is as relevant as the price estimate. All three models are distributed under the NVIDIA OneWay Noncommercial License, which permits research or evaluation and applies the same restriction to derivatives. That makes Flora a potential research tool and comparison point, but not a freely usable commercial labeling service under the stated terms. Nvidia’s own commercial use does not change the license granted to other users.
The reported savings also do not establish total project economics. The headline comparison concerns captioning on a small fixed test set; it does not include the reward model or all infrastructure costs. Lower labeling expense would address one part of robot development, while collecting useful video, building and maintaining hardware, and validating safe behavior remain separate challenges.
Products Worth Considering
Bose QuietComfort Ultra Earbuds (2nd Gen), Deep Plum
Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Eucalyptus Green
Apple iPad Air 13-inch (M4), 128GB, Wi-Fi 7, Space Gray
As an affiliate, we earn on qualifying purchases.
How Nvidia Built the Comparison
Robot-learning systems use video from both robots and people, including recordings of tasks such as moving objects or folding fabric. The Flora paper describes three kinds of supervision: timestamped captions tied to actions, step-level process rewards that distinguish normal, suboptimal and failed behavior, and overall outcome scores. These labels can help developers organize training data and evaluate attempts.
The test set contained 20 human-hand episodes, 15 episodes with two-finger grippers and 15 with dexterous robot hands. Nvidia compared Flora with GPT-6 Astra and other models on its benchmark. The cost appendix states that 106 of 7,062 Astra attempts reported no usage, so the reported API total is a known-usage subtotal. Nvidia’s paper also says its production GPU-hour denominator has not been verified.
The models are substantial: the captioner and rewarder each have 27 billion parameters. The source material estimates about 55 GB of model weights per 27-billion-parameter model at 16-bit precision, placing deployment beyond the typical gaming PC. This matters when comparing a GPU-cost estimate with a hosted API bill: the hardware, operating costs and workload assumptions affect what a user would actually pay.
“research or evaluation purposes only”
— NVIDIA OneWay Noncommercial License
What the Benchmark Cannot Show
The cost comparison does not establish the price of running Flora in a commercial production environment. Its GPU rate and throughput are assumptions, the production GPU-hour denominator is unverified, and the estimate excludes storage and overhead. The Astra figure, by contrast, reflects reported API usage and has a known-usage caveat of its own. Nvidia’s paper does not provide an independently verified, like-for-like operating-cost audit.
The benchmark also covers about 75 minutes of selected video, not a broad sample of real-world deployments. Nvidia ran most tests itself, and the reported scores show different strengths across video types and tasks. It remains unclear how the results would hold across other datasets, longer workloads, different hardware or independent evaluations. The 125× figure applies to captioning costs as presented; it should not be read as a cost reduction for every Flora function or an entire robotics program.
The supplied material does not establish whether Nvidia will offer a separate commercial license, revise the license terms or provide a hosted commercial service. Those details would affect whether companies can use Flora beyond research and evaluation.
License and Deployment Questions
Researchers can inspect the paper, model cards and model repositories to test Flora against their own datasets, subject to the license. The next useful checks are independent replications of the benchmark and measured deployments that report hardware utilization, throughput, storage and operating costs. Such results could show whether the modeled savings carry over to workloads beyond Nvidia’s test episodes.
Companies considering the models will need to review the noncommercial terms before using them in products or paid services. Any commercial deployment would require terms that permit it, if Nvidia offers them. Further details about licensing, independent performance tests and real-world operating costs have not been announced in the source material.
Key Questions
What is Nvidia Flora?
Flora is a three-model suite for video labeling in robotics: one model finds action intervals, one captions them, and one scores robot performance.
Is Flora really 125 times cheaper than GPT-6 Astra?
Nvidia reports a ratio of about 122.5 to 1, rounded to 125×, for captioning on 50 test episodes. Flora’s $0.72 figure is modeled using assumed GPU costs; Astra’s $88.49 figure is based on API usage, so the comparison is not like-for-like.
Can a company use Flora commercially?
The models are distributed under the NVIDIA OneWay Noncommercial License, which the source describes as allowing research or evaluation only. The supplied material does not say whether Nvidia offers separate commercial terms.
Does Flora outperform GPT-6 Astra on every task?
No. Nvidia’s reported overall score is 57.1% for Flora versus 53.7% for Astra, but the paper reports different winners by task: Flora does better on human videos and partial-success cases, while Astra leads on robot videos, caption accuracy and rubric scoring.
What is still unknown about Flora’s cost?
The production GPU-hour denominator in Nvidia’s model has not been verified, and the estimate excludes storage and overhead. Independent measurements across larger or different workloads are not provided in the source material.
Source: ThorstenMeyerAI.com
