🔍 Read the full analysis: Nvidia’s Flora: Lower-Cost Robot Video Labeling, Plus A License Catch 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 video. Nvidia estimates captioning costs $0.72 per video hour, versus $88.49 for GPT-6 Astra on a small benchmark, but Flora’s figure is modeled, its quality results are mixed, and its license restricts use to research or evaluation.
Nvidia’s GEAR robotics lab has released Flora, a set of three models designed to segment, describe and score robot video, and says its captioning system could cost $0.72 per video hour. That figure is a modeled GPU estimate, not a measured production bill, and Flora’s license permits research or evaluation use only, not commercial deployment.
Flora consists of Flora-TempSeg, a 9-billion-parameter model that identifies action boundaries; Flora-Captioner, a 27-billion-parameter model that writes instructions for each segment; and Flora-Rewarder, a 27-billion-parameter model that scores an attempt or grades its steps. The models are based on Alibaba’s Qwen3.5 family. Nvidia says Flora was used as a data engine for its GR00T-2 robot model.
The headline cost comparison concerns captioning. Nvidia’s paper reports $88.49 per video hour for GPT-6 Astra API usage on 50 episodes, about 75 minutes of footage, and estimates Flora at $0.72 per hour using eight GB300 GPUs priced at an assumed $10 per GPU-hour. Dividing those figures gives about 122.5, rounded by Nvidia to approximately 125 times cheaper. The paper says the production GPU-hour denominator has not been verified; the estimate excludes storage and overhead and gives a range of $0.51 to $1.26.
On the same test set, Nvidia reports Flora scored 57.1%, compared with 53.7% for GPT-6 Astra. The broader results are not uniformly in Flora’s favor: it performed well on human videos, action boundaries and partial-success cases, while Astra led on robot videos, caption accuracy and rubric scoring. The cost claim covers captioning, not the separate 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, With Limits
Video annotation is one step in turning demonstrations into training data for robots. Models that identify actions, describe them and flag failed or incomplete steps can help researchers sort large collections of footage and supply feedback for training. If Flora’s estimated costs hold at scale, lower-cost labeling could reduce one expense in building robot datasets.
But the reported saving should not be treated as a verified operating cost or proof that companies can use Flora in commercial products. The comparison is between measured API usage and a modeled GPU budget, based on a limited test set selected and evaluated by Nvidia. The noncommercial license also means businesses cannot simply adopt the released models as a free production service. They can serve as a research reference, but commercial use would require a different permission or an independently developed alternative.
Cheaper labels would not by themselves solve other barriers to robotics. Collecting representative robot data, building reliable hardware and demonstrating safety remain separate challenges. Flora addresses the annotation task, not the full cost or difficulty of deploying useful robots.
Products Worth Considering
Logitech MX Mechanical Full-Size Wireless Mechanical Keyboard, Tactile Quiet Switches, Backlit Keys, Bluetooth, USB-C, Multi-Device, Graphit
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
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 video of people and machines performing tasks, such as placing an object or folding fabric. As described in the Flora paper, training workflows can use timestamped captions to identify what happened and when, step-level rewards to mark normal, suboptimal or failed actions, and scores for an entire attempt. These labels help researchers organize training examples and assess performance.
Flora divides those jobs among its three models. TempSeg finds intervals in a video; Captioner writes an English instruction for each interval, using surrounding footage for context; Rewarder can grade half-second steps or score an episode against a user’s rubric. The models are available through Hugging Face repositories, with inference code included there. Nvidia’s model cards say they were trained on fewer than 10,000 hours of human and robot video.
The benchmark described in the paper used 50 episodes totaling 4,483 seconds: 20 filmed with human hands, 15 with two-finger grippers and 15 with dexterous robot hands. Nvidia also reports that Flora can label up to 200,000 hours of video per day on 512 GB300 GPUs. That throughput is Nvidia’s claim, not an independently verified production result.
“Research or evaluation purposes only.”
— NVIDIA OneWay Noncommercial License
Cost and Quality Still Need Testing
The $0.72 figure is not a measured production price. Nvidia modeled it using an assumed $10 per GPU-hour and projected throughput, while the GPT-6 Astra figure comes from API usage on the test episodes. Nvidia’s paper says the GPU-hour denominator has not been verified, and the calculation excludes storage and overhead. The 50-episode test is also small, and the reported results come from Nvidia’s evaluation rather than an independent replication.
Flora’s benchmark performance varies by task and video type, so the overall score does not show that it is better for every labeling job. The provided material does not establish how the models will perform across other datasets, camera setups or real-world production workloads. It also does not state whether Nvidia will offer a commercial license or another paid route for companies seeking to use Flora.
Independent Tests and License Terms
The next useful test is whether other researchers can reproduce Nvidia’s cost and quality results using different robot datasets and clearly reported hardware, throughput and operating expenses. Production measurements would show how much the modeled estimate changes when storage, system overhead and workload variability are included.
For organizations considering the models, the immediate step is to review the NVIDIA OneWay Noncommercial License before use. Nvidia has released the models and inference code through Hugging Face, but the supplied information does not announce a commercial-use option or a date for further benchmark updates. Those details remain open.
Key Questions
What is Nvidia Flora?
Flora is a set of three video models from Nvidia’s GEAR robotics lab. They identify action segments, describe those segments in language and score 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 an assumed $10 per GPU-hour and projected throughput. The paper says the production GPU-hour denominator has not been verified, and the estimate excludes storage and overhead.
How did Nvidia calculate the approximately 125-times comparison?
Nvidia compared its estimated $0.72 per video hour for Flora captioning with $88.49 per hour from GPT-6 Astra API usage on 50 episodes. The ratio is about 122.5, rounded to approximately 125. The two figures are based on different cost methods: a modeled GPU budget and reported API usage.
Can companies use Flora commercially?
The models are released under the NVIDIA OneWay Noncommercial License, which limits use to research or evaluation. The supplied material does not provide commercial-use rights; companies would need to check the license and obtain any required separate permission.
Did Flora outperform GPT-6 Astra in every test?
No. Nvidia reports Flora’s overall score as 57.1% versus Astra’s 53.7% on the specified test set, but the results vary by category. Flora did better on some human-video, action-boundary and partial-success tests, while Astra led on robot videos, caption accuracy and rubric scoring.
Source: ThorstenMeyerAI.com
