🔍 Read the full analysis: Robot Video Labeling At 125× Lower Cost: Nvidia Flora’s Promise And Caveat on ThorstenMeyerAI.com
Get the latest gadgets 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 robotics lab has released Flora, three models for segmenting, describing and scoring robot video. Nvidia estimates captioning costs 72 cents per video hour, about 122.5 times below a GPT-6 Astra API comparison, but Flora’s cost is modeled, its quality results are mixed, and its license restricts use to research and evaluation.
Nvidia’s GEAR robotics lab has released Flora, a three-model system for labeling robot video, and estimates that its captioning can cost $0.72 per video hour compared with $88.49 using GPT-6 Astra on a selected test set. The implied ratio is about 122.5 times, rounded by Nvidia to approximately 125 times, but the Flora figure is a modeled GPU cost rather than a measured production bill, and the models’ license limits use to research or evaluation.
Flora consists 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 an episode or grades short 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 and was trained on fewer than 10,000 hours of human and robot video.
The price comparison uses 50 episodes from Nvidia’s DenseCaptionBench, totaling 4,483 seconds, or about 75 minutes. Nvidia reports $88.49 per video hour for GPT-6 Astra API usage, while Flora’s $0.72 estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput. The paper gives Flora scenarios from $0.51 to $1.26 per hour and says the production GPU-hour denominator has not been verified. The estimate excludes storage and overhead; the 125-times comparison covers captioning, not Flora’s reward model.
On that same test set, Nvidia reports a quality score of 57.1% for Flora and 53.7% for Astra. The paper’s other benchmark results are not uniformly in Flora’s favor: it reports wins for Flora on human videos, action boundaries and partial-success evaluation, while Astra does better on robot videos, caption accuracy and rubric scoring. Nvidia conducted most of the evaluations, so the results should be read as vendor-reported benchmarks rather than independent confirmation.
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 Label Costs, Limited Deployment
Video labels describe what happened in a recording, when it happened and whether a robot’s action succeeded. That information helps researchers filter training data, identify failures and assess behavior. If Flora’s cost advantage holds at scale, it could reduce one expense involved in building and evaluating robot systems, particularly as training sets grow.
But the headline number is not a direct comparison of two production invoices. One side is a measured API usage total; the other is a GPU-cost model with an assumed rental rate and throughput. Nvidia’s own scenarios still put Flora below Astra’s reported price, but the result does not establish the cost of running a full commercial data pipeline. The license also means businesses cannot treat the released models as a free production labeling service.
Cheaper labeling would address only one part of robotics development. Collecting useful robot data, operating hardware, testing systems safely and turning benchmark gains into dependable performance remain separate challenges. Flora may offer researchers a lower-cost tool and companies a reference point, but the provided results do not show that it resolves those wider constraints.
Products Worth Considering
Logitech MX Keys S Wireless Keyboard, Graphite
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
Samsung 32-Inch Odyssey OLED G8 (G80SH) 4K Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Three Models for Video Labels
Robot-learning datasets contain recordings of actions such as moving an object, handling a tool or folding fabric. The Flora paper describes three labeling needs: timestamped captions that connect instructions to moments in a clip, process rewards that grade individual steps, and outcome scores that assess a full attempt. Flora assigns those tasks across its segmenter, captioner and rewarder.
Nvidia says the models were trained on fewer than 10,000 hours of human and robot video and were used in the development of GR00T-2. The model cards list inference code in the models’ Hugging Face repositories. The source material estimates that running either 27-billion-parameter model at 16-bit precision would require roughly 55 GB for weights, placing deployment closer to data-center hardware than a typical consumer computer.
The release is not an unrestricted commercial offering. All three models use the NVIDIA OneWay Noncommercial License, which permits research or evaluation and applies the same limitation to derivatives. The source material says Nvidia may use the models commercially; prospective users should check the license itself for the terms that apply to their work.
“research or evaluation purposes only”
— NVIDIA OneWay Noncommercial License
Cost and Quality Still Need Testing
The cost estimate depends on an assumed $10 GPU-hour rate, modeled throughput and a production denominator Nvidia says has not been verified. The comparison also uses a small, vendor-selected set of 50 episodes, and the Astra API total includes only attempts with reported usage; 106 of 7,062 attempts had no usage reported. The available material does not establish how results would change across different datasets, workloads or operating conditions.
Quality results vary by task and video type. Nvidia reports Flora ahead on some measures and Astra ahead on others, but independent replication is not provided in the source material. It is also unclear how much the reported costs would change once storage, operational overhead and the full costs of deployment are included. The $0.72 figure should not be interpreted as a complete price for labeling every kind of robot data.
Independent Benchmarks and License Checks
The next useful evidence would be independent testing on larger and more varied robot-video datasets, alongside transparent reporting of the compute used and the full cost of running the models. Such tests could show whether Flora’s reported advantages on human footage and partial successes carry over beyond Nvidia’s benchmarks, and how its weaker results on some robot-video and scoring tasks affect practical use.
Researchers and companies considering the release should also review the license before adopting it or modifying the models. Nvidia’s announced results do not specify a timeline for broader evaluations or a commercial license. Until those details emerge, the clearest conclusion is limited: Flora offers a set of research-oriented video-labeling models and a potentially large modeled cost reduction, not a verified commercial price or a uniformly stronger alternative.
Key Questions
What is Nvidia Flora?
Flora is a set of three models for segmenting robot-video actions, writing captions for them and scoring robot performance. Nvidia’s GEAR lab says it was used as a data engine for GR00T-2.
Is Flora really 125 times cheaper than GPT-6 Astra?
Nvidia’s comparison works out to about 122.5 times, rounded to approximately 125 times. It compares Astra’s reported API usage with a modeled Flora GPU cost of $0.72 per video hour, not two measured production bills.
Does Flora outperform GPT-6 Astra?
Not across every reported task. Nvidia reports a higher score for Flora on the selected test set overall and advantages on some benchmarks, including human-video evaluation. Astra performs better on some robot-video, caption-accuracy and rubric-scoring measures.
Can a company use Flora commercially?
The models are released under the NVIDIA OneWay Noncommercial License, described as allowing research or evaluation purposes only. Companies should review the license’s actual terms before use; the release is not presented as an unrestricted commercial pipeline.
What costs are excluded from Nvidia’s estimate?
The $0.72-per-hour estimate is based on assumed GPU rental pricing and modeled throughput. The source material says it excludes storage and overhead, and Nvidia notes that the production GPU-hour denominator has not been verified.
Source: ThorstenMeyerAI.com
