🔍 Read the full analysis: Why Nvidia’s Flora Is Drawing Attention In Robot Video Labeling 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, three models for segmenting, captioning and scoring robot video. Nvidia estimates captioning costs $0.72 per video hour, compared with $88.49 for GPT-6 Astra, but Flora’s cost is modeled rather than measured, its benchmark 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 mark actions in video, describe them and score robot performance. Nvidia says captioning with Flora costs a modeled $0.72 per video hour, versus $88.49 for GPT-6 Astra on its test, a comparison drawing attention to the cost of preparing data for robot training. The estimate is not a measured production bill, and Flora’s license permits research or evaluation use only.
Flora comprises three models built on Alibaba’s Qwen3.5. Flora-TempSeg, with 9 billion parameters, identifies the time boundaries of actions. The 27-billion-parameter Flora-Captioner writes an instruction for each segment, while the 27-billion-parameter Flora-Rewarder scores an entire episode, grades short steps and can apply a user-provided rubric. Nvidia says the system was used as a data engine for its GR00T-2 robot models.
The cost comparison in the paper uses 50 episodes totaling 4,483 seconds—about 75 minutes—from Nvidia’s DenseCaptionBench. GPT-6 Astra’s $88.49 per video hour is based on reported API usage at published rates. Flora’s $0.72 figure is instead a model: it assumes eight GB300 GPUs rented at $10 each per hour and estimated throughput. Nvidia’s paper gives a range of $0.51 to $1.26 per video hour, excluding storage and other overhead, and says the production GPU-hour denominator has not been verified. The comparison covers captioning, not Flora’s reward model.
On the same test set, Nvidia reports a quality score of 57.1% for Flora and 53.7% for Astra. The paper’s other results are mixed: Flora performed better on human video, action boundaries and some partial-success assessments, while Astra scored higher on robot-video caption accuracy and rubric scoring. Nvidia conducted most of the testing, so the results are not an independent assessment.
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
Robot-learning systems need annotated examples showing what action occurred, when it happened and whether it succeeded. Producing those labels can be a costly step in preparing video for training and evaluation. If Flora’s estimated throughput and GPU costs hold up in practical deployments, lower-cost labeling could let robotics teams process more footage or reduce their reliance on per-use commercial APIs.
The result is not, on its own, evidence that robot development becomes inexpensive. Nvidia’s comparison is confined to a small set of episodes and a modeled hardware budget; collecting useful robot video, running and maintaining hardware, and testing systems safely remain separate expenses. The noncommercial license also means companies cannot treat the released models as a free production service. For commercial teams, Flora may offer a reference for evaluation or a starting point for research, subject to the license, rather than a drop-in business tool.
NVIDIA GPU workstation for AI model inference
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Three Models, Three Labeling Tasks
Robot video can show a gripper placing a mug, two hands passing a part or a person completing a household task. Training teams use timestamped captions to find and describe actions in those recordings. They can also use step-level rewards to identify failures and episode-level scores to track overall performance. Flora packages these tasks in separate models rather than presenting a single captioning system as the whole labeling pipeline.
According to the paper and model cards, Flora was trained on fewer than 10,000 hours of human and robot video. Nvidia says it can label as much as 200,000 hours per day using 512 GB300 GPUs; that is a company-reported capacity figure, not a result established by the 50-episode cost test. The paper’s estimate compares a modeled cluster cost with API charges, so the reported 122.5-to-1 ratio—rounded by Nvidia to about 125 times—should be read with that distinction in mind.
The models are distributed through Hugging Face repositories under the NVIDIA OneWay Noncommercial License. Its stated scope is research or evaluation, and derivatives inherit the restriction. Nvidia may use the models commercially under its own rights, but that does not grant commercial rights to outside users.
“Research or evaluation purposes only.”
— NVIDIA OneWay Noncommercial License
NVIDIA DGX workstation for robotics research
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Cost and Quality Need Wider Tests
It remains unclear how Flora’s modeled price would translate to a sustained production workload. Nvidia’s estimate assumes a specific GPU rental rate and modeled throughput; the paper says the production GPU-hour denominator has not been verified. Its Astra figure also relies on reported usage for the test, and the comparison uses about 75 minutes of video selected for Nvidia’s benchmark, not a broad, independently run evaluation.
The quality results do not establish that Flora is better across robot tasks. It scored higher on the reported overall test and several subsets, but Astra led on robot-video caption accuracy and rubric scoring. The source material does not provide an independent replication or show how results generalize to different robots, camera setups and real-world operating conditions. Commercial deployment is separately unresolved for outside users because the stated license is noncommercial.
robotics development kit with camera and sensors
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Independent Replication Will Matter
The next useful evidence would be independent tests on larger and more varied video sets, with clear reporting of actual GPU hours, hardware utilization, storage and operating costs. Such work could show whether the projected price holds beyond Nvidia’s chosen episodes and whether the models’ different strengths persist across robot types and tasks.
Organizations considering Flora will also need to review the license before use and distinguish research evaluation from commercial deployment. Nvidia says Flora supports its GR00T data work, but the release information does not set out a schedule for broader access, a commercial license for outside users or further validation. Those details remain open.
high speed NVMe SSD for AI video 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 suite of three video models from Nvidia’s GEAR robotics lab. They segment actions, caption those segments and score robot performance.
Is the $0.72-per-hour figure a measured cost?
No. Nvidia’s $0.72 per video hour is modeled using assumed GPU rental prices and estimated throughput. The paper gives a range of $0.51 to $1.26 and says the production GPU-hour denominator has not been verified.
Does Flora beat GPT-6 Astra on every test?
No. Nvidia reports Flora ahead on the overall score and on some human-video, segmentation and partial-success results. Astra scored higher on robot-video caption accuracy and rubric scoring in the reported tests.
Can a company use Flora in a commercial product?
The models are released under a license limited to research or evaluation purposes, with derivatives subject to the same restriction. Companies should not assume the release permits commercial use.
How much robot video did Nvidia use for its cost comparison?
The comparison used 50 episodes totaling 4,483 seconds, or about 75 minutes, from Nvidia’s DenseCaptionBench. The cost result has not been established on a large, independently tested production workload.
Source: ThorstenMeyerAI.com
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
