🔍 Read the full analysis: What’s Behind Nvidia Flora’s 125× Cheaper Robot Video Labeling? on ThorstenMeyerAI.com
Get the little things that make your day 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 videos. Nvidia estimates captioning costs 72 cents per video hour, about 122.5 times below a GPT-6 Astra API benchmark, but the comparison pairs a modeled GPU cost with measured API usage and uses a small Nvidia-selected test set. Flora’s license permits research and evaluation, not commercial use.
Nvidia’s GEAR robotics lab has released Flora, a set of three models that segment, caption and score robot videos, and estimates that its captioning pipeline costs $0.72 per video hour. Nvidia compares that modeled GPU cost with $88.49 per hour of measured API usage for GPT-6 Astra, describing the difference as about 125 times; the figures do not come from the same kind of cost measurement, and Flora’s license restricts use to research and evaluation.
The three models perform separate tasks. Flora-TempSeg, with 9 billion parameters, identifies the time intervals in which actions occur. The 27-billion-parameter Flora-Captioner writes an instruction for each interval, while the 27-billion-parameter Flora-Rewarder can grade video episodes and individual steps. Nvidia says the models were built on Alibaba’s Qwen3.5 models and that Flora served as a data engine for its GR00T-2 robot models.
Nvidia’s cost comparison uses 50 episodes totaling 4,483 seconds—about 75 minutes—from its DenseCaptionBench. The paper reports GPT-6 Astra API usage at $88.49 per video hour and estimates Flora at $0.72, based on eight GB300 GPUs priced at an assumed $10 per GPU-hour and modeled throughput. Dividing those amounts gives about 122.5 times, which Nvidia rounds to approximately 125 times. The estimate excludes storage and other overhead; the paper says its production GPU-hour denominator has not been verified.
On that same test set, the source material reports an overall Flora score of 57.1%, compared with 53.7% for Astra. The results vary by task: Flora performs better on human-video examples, action boundaries and partial successes, while Astra leads on robot-video examples, caption accuracy and rubric scoring. The cost claim applies to captioning; it does not include the 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 Costs Could Expand Robot Data
Video labels help robot developers connect recorded actions with instructions and judgments about whether a step succeeded. If a labeling system can process large collections at substantially lower cost, teams may be able to annotate more examples or run more experiments. That could ease one part of building robot-training datasets, although it does not itself solve the collection of useful robot footage, hardware expense or safety testing.
The comparison is relevant, but not a direct like-for-like production test. Flora’s figure is a modeled GPU budget using an assumed rental price, while Astra’s is reported API usage. Nvidia’s test was also limited to a fixed set of 50 episodes selected for its benchmark. The result offers a cost estimate and a benchmark comparison, not proof that all companies will obtain the same savings at scale.
Commercial access is another limit. The models are released under the NVIDIA OneWay Noncommercial License, which allows research or evaluation and applies the same restriction to derivatives. That makes Flora a potential research tool and comparison point, but not a no-cost commercial labeling service for companies.
Products Worth Considering
Oura Ring 4 Smart Ring – Silver – Size 8
BowFlex Results Series SelectTech 552 Adjustable Dumbbell
Philips Sonicare 5950 Rechargeable Electric Toothbrush with 2 Brush Heads, Navy
As an affiliate, we earn on qualifying purchases.
Three Models, Three Labeling Tasks
Robot-learning datasets can include first-person human demonstrations, footage from robot grippers and recordings of more dexterous hands. The Flora paper describes three kinds of labels: timestamped captions tied to actions, step-level process rewards that mark behavior as normal, suboptimal or failed, and an overall outcome score for an attempt. These labels can help developers sort training data and identify where a robot’s attempt went wrong.
Flora splits those jobs across its segmenter, captioner and rewarder, rather than treating video annotation as one task. According to the source material, the models were trained on fewer than 10,000 hours of human and robot video. Nvidia says the project is intended to support its GR00T robotics work. The models are available through Hugging Face repositories, with inference code included there.
The hardware requirement may limit who can run the larger models. The source estimates that each 27-billion-parameter model needs roughly 55 GB for weights at 16-bit precision, placing local inference in data-center-class territory rather than on an ordinary gaming PC. Nvidia’s tests used an H100 GPU on Linux; the reported low price estimate instead assumes access to GB300 GPUs at a specified hourly rate.
“research or evaluation purposes only”
— NVIDIA OneWay Noncommercial License
Cost and Benchmark Limits
The headline ratio depends on an assumed GPU price and modeled throughput; the paper says its production GPU-hour denominator has not been verified. It also compares that estimate with reported API usage, rather than with a measured Flora bill under the same operating conditions. Nvidia gives cost scenarios from $0.51 to $1.26 per video hour, excluding storage and overhead.
The quality results are not uniform across tasks, and the evaluation covered 50 episodes in one benchmark. The supplied material does not establish whether the scores or costs would hold across other datasets, longer production runs or different hardware and operating arrangements. It also does not provide enough information to assess independent replication of the reported results. The 125-times comparison concerns captioning, not all three Flora models or every cost involved in running a labeling pipeline.
Commercial Use Remains Restricted
Researchers can assess Flora under its stated research and evaluation terms, and developers can compare its outputs with other annotation systems on their own video. Companies considering deployment will need to review the license and cannot treat the release as unrestricted commercial software. The source material does not identify a date for a broader license, additional benchmark results or a verified production-cost measurement.
The next useful evidence would be results from independent evaluations, broader video sets and production runs that report actual GPU utilization, throughput and overhead. Until then, Nvidia’s figures support a promising cost estimate—not a guarantee of a particular saving or labeling quality for other users.
Key Questions
What is Nvidia Flora?
Flora is a set of three video models from Nvidia’s GEAR robotics lab. They identify action intervals, write captions for those intervals and score robot behavior.
How did Nvidia calculate the approximately 125× cost difference?
Nvidia compared $88.49 per video hour in reported GPT-6 Astra API usage with a Flora cost estimate of $0.72 per hour based on eight GB300 GPUs at an assumed $10 per GPU-hour. The ratio is about 122.5, rounded to approximately 125. The Flora amount is modeled, not a measured production bill.
Does Flora outperform GPT-6 Astra on every test?
No. The reported overall scores were 57.1% for Flora and 53.7% for Astra on the 50-episode test set, but results differed by task. Flora did better on human videos, action boundaries and partial successes; Astra led on robot videos, caption accuracy and rubric scoring.
Can a company use Flora in a commercial product?
The models are released under the NVIDIA OneWay Noncommercial License, which the source describes as allowing research or evaluation only. Derivatives inherit that restriction, so the release is not an unrestricted commercial pipeline.
What remains unknown about the savings?
It is not yet clear whether the modeled cost will hold in production across different datasets, hardware setups and operating conditions. Nvidia’s paper says the production GPU-hour denominator has not been verified, and its estimate excludes storage and other overhead.
Source: ThorstenMeyerAI.com
