🔍 Read the full analysis: Nvidia’s Flora Targets Cheaper Robot Video Labeling, But The License Is Part Of The Story 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 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, quality results are mixed, and the models are licensed for research or evaluation only.
Nvidia’s GEAR robotics lab has released Flora, three models designed to label video used in robot training, and estimates that one part of the system can caption video for $0.72 per hour. That figure is a modeled GPU cost, not a measured production bill; the same comparison puts GPT-6 Astra API usage at $88.49 per hour, while Flora’s license permits research or evaluation only, not commercial use.
Flora comprises three models built on Alibaba’s Qwen3.5: Flora-TempSeg, a 9-billion-parameter model that identifies the start and end of actions; Flora-Captioner, a 27-billion-parameter model that writes instructions for each segment; and Flora-Rewarder, also 27 billion parameters, which scores an episode or grades short steps as normal, suboptimal or failure. The models are available through Hugging Face repositories, according to the source material.
Nvidia’s paper compares captioning costs on 50 episodes totaling 4,483 seconds, or about 75 minutes. It reports $88.49 per video hour for GPT-6 Astra, based on API usage, and estimates Flora at $0.72 using eight GB300 GPUs priced at an assumed $10 per GPU-hour. The paper gives Flora cost scenarios from $0.51 to $1.26 per hour and says the production GPU-hour denominator has not been verified. The estimated ratio is 122.5 to one, rounded by Nvidia to approximately 125 times.
The cost calculation covers captioning only; it does not include Flora’s reward model. Nvidia reports Flora scored 57.1% on its quality comparison, against 53.7% for Astra, 38.3% for Gemini 3.8 Flash and 30.4% for Claude Opus 5. The source material describes the results as mixed: Flora performed better on human videos, action boundaries and partial-success judgments, while Astra led on robot-video caption accuracy and rubric scoring. Nvidia conducted most of the tests itself.
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 labels tell a robot-learning system what happened, when it happened and whether an action worked. Producing those labels at scale can be a significant cost and labor burden for developers assembling training data. A lower-cost labeling tool could make it easier to sort and evaluate larger video collections, including footage of failed or incomplete actions.
But Flora’s headline cost estimate should not be read as a verified commercial price. It compares a modeled GPU budget with an API bill measured on a small, fixed test set selected by Nvidia. The license also limits the models’ immediate use by businesses: companies cannot treat the release as a free production pipeline under the stated terms. The results may still offer a reference point for researchers and a design example for firms considering their own systems.
Cheaper annotation would address only one part of robot development. The source material points to data collection, safety and hardware costs as continuing constraints. The release does not establish that lower labeling costs alone will make household or service robots practical or less expensive.
NVIDIA RTX workstation GPU for AI model inference
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Three Models for Robot Videos
Robot foundation models learn from recordings of tasks such as placing a mug, handling a wire or folding a towel. Nvidia’s paper describes three kinds of labels used with those recordings: timestamped captions tied to video intervals, process rewards for individual steps, and outcome scores for an entire attempt. These can help researchers organize training data and assess robot behavior.
According to the model cards cited in the source material, Flora served as a data engine for Nvidia’s GR00T-2 and was trained on fewer than 10,000 hours of human and robot video. Nvidia says its system can label as much as 200,000 hours of video per day using 512 GB300 GPUs. That is a company-reported capacity figure, not an independently verified production measurement.
The models are distributed under the NVIDIA OneWay Noncommercial License. The source material says it restricts use to research or evaluation and requires derivatives to retain the same limitation. It also reports that Nvidia may use the models commercially. The comparison and license terms therefore matter alongside the code and model availability when organizations decide whether Flora fits their plans.
“approximately 125×”
— Nvidia’s paper, describing its cost comparison
GPU server with multiple NVIDIA GPUs for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs and Results Need Validation
The Flora cost figure remains an estimate based on assumed GPU pricing and modeled throughput. Nvidia’s paper states that the production GPU-hour denominator has not been verified, and the comparison excludes storage and other overhead. The test covered roughly 75 minutes of video in 50 episodes, so the reported figures do not establish what labeling would cost across other datasets, hardware arrangements or production workloads.
The quality comparison also has limits: the source material says Nvidia ran most tests, and Flora’s reported advantage is not consistent across tasks. Astra scored better on robot videos, caption accuracy and rubric scoring, while Flora performed better on some human-video and partial-success measures. It is not clear from the supplied material how independent testing or broader real-world use would change those results.
The reported daily throughput of up to 200,000 video hours on 512 GB300 GPUs is Nvidia’s claim. The material does not provide independent verification of that capacity under commercial operating conditions. Nor does it establish whether Nvidia will offer different licensing terms or a commercial version.
robotics development kit with camera and sensors
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Tests and License Terms
The next useful evidence will be independent evaluations across a wider mix of human and robot videos, with costs measured on production workloads rather than modeled from assumed GPU rates. Researchers and developers will also need to compare performance across the separate segmentation, captioning and reward tasks; the 125-times estimate applies only to captioning.
For organizations considering deployment, the immediate next step is to review the license and model cards before building a workflow around Flora. The source material does not identify a scheduled commercial release, a change to the noncommercial terms or a date for further benchmark results. Until those details emerge, the release is most directly relevant to research and evaluation rather than commercial production.
high-speed NVMe SSD for AI training datasets
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Nvidia Flora?
Flora is a set of three video models from Nvidia’s GEAR robotics lab: one segments video into action intervals, one captions those intervals, and one scores robot performance. Nvidia says the models are intended to support labeling video used in robot training.
Is Flora proven to cost $0.72 per hour?
No. $0.72 per video hour is a modeled estimate based on eight GB300 GPUs, an assumed rate of $10 per GPU-hour and modeled throughput. Nvidia’s paper says the production GPU-hour denominator has not been verified.
Does the 125-times comparison cover all of Flora?
No. The comparison is for captioning only; it does not include the Flora-Rewarder. It compares a modeled Flora GPU cost with measured GPT-6 Astra API usage on 50 episodes totaling about 75 minutes.
Can a company use Flora commercially?
According to the source material, the models are released under the NVIDIA OneWay Noncommercial License, which allows research or evaluation use only. The material says derivatives inherit that restriction; companies should consult the license itself for the applicable terms.
Does Flora outperform GPT-6 Astra on every test?
No. Nvidia’s reported results show Flora ahead on some measures, including human videos and partial-success judgments, while Astra leads on robot-video caption accuracy and rubric scoring. Most tests were run by Nvidia, according to the source material.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
