Why Nvidia’s Flora Makes Robot Video Labeling 125× Cheaper, With A License Catch

🔍 Read the full analysis: Why Nvidia’s Flora Makes Robot Video Labeling 125× Cheaper, With A License Catch on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Nvidia’s GEAR robotics lab has released Flora, a suite of three models designed to segment, caption and score robot videos, and estimates its captioning costs at $0.72 per video hour—about 122.5 times below a GPT-6 Astra comparison. The cost figure is a modeled GPU estimate, not a measured production bill, and the models’ noncommercial license bars business use.

Flora comprises Flora-TempSeg, a 9-billion-parameter model that marks action boundaries; Flora-Captioner, a 27-billion-parameter model that writes instructions for each segment; and Flora-Rewarder, another 27-billion-parameter model that rates an episode or grades steps. Nvidia says the system was used as a data engine for its GR00T-2 robot models. The models are built from Alibaba’s Qwen3.5 family and are available through Hugging Face repositories.

The cost comparison in Nvidia’s paper uses 50 episodes totaling 4,483 seconds, or about 75 minutes, from Nvidia’s DenseCaptionBench. GPT-6 Astra’s reported $88.49 per video hour reflects API usage at public rates cited for October 4, 2026. Flora’s $0.72 figure is estimated from modeled throughput on eight GB300 GPUs, priced at an assumed $10 per GPU-hour. Nvidia gives a range of $0.51 to $1.26 and says the calculation excludes storage and overhead.

On the same episodes, Nvidia reports 57.1% for Flora and 53.7% for Astra, with lower scores for Gemini 3.8 Flash and Claude Opus 5. The comparison is not a broad independent evaluation: Nvidia ran most tests, and results vary by task. Flora performs better on some human-video and partial-success measures, while Astra leads on robot-video caption accuracy and rubric scoring. The 125× cost claim applies to captioning only, not Flora’s reward model.

Lower Costs, Limited Business Access

Video annotation is a costly step in building robot training datasets: a system must identify when actions occur, describe them and assess whether they succeeded. If automated labeling works at scale, it could reduce the expense and time involved in preparing examples for robot foundation models. Nvidia says Flora can process as much as 200,000 hours of video per day on 512 GB300 GPUs, a company-reported capacity rather than an independently verified production result.

The license limits who can use that potential directly. All three models are distributed under the NVIDIA OneWay Noncommercial License, which permits research or evaluation, with derivatives subject to the same restriction. That makes Flora relevant to researchers and companies testing labeling methods, but not a free commercial pipeline. Businesses would need separate permission or another tool for production use.

The cost ratio also needs careful reading. It compares a modeled GPU budget with API charges observed on a small, vendor-selected set of episodes. Nvidia’s paper says the production GPU-hour denominator has not been verified. The result suggests a possible cost advantage, but does not establish what a company would pay after hardware, engineering, storage and operating costs.

How Flora Labels Robot Footage

Robot-learning datasets can include recordings of robotic grippers and human hands performing tasks such as placing a mug or folding a towel. According to the Flora paper, training and monitoring systems can use timestamped captions to locate and describe actions, step-level rewards to identify failures or suboptimal behavior, and overall outcome scores to assess a full attempt.

Flora’s three models divide those functions. TempSeg proposes intervals where actions begin and end; Captioner describes each interval in English and can use a main camera and wrist-camera footage; Rewarder gives an episode a score from 1 to 5, labels half-second steps as normal, suboptimal or failure, or evaluates against a supplied rubric. Nvidia says the models were trained on fewer than 10,000 hours of human and robot video.

The paper’s benchmark results do not point to a universal winner. Flora scored 50.3 against Astra’s 46.0 on the public WGO-Bench segmentation measure, and 37.4 against 29.7 on Nvidia’s human-video subset. On robot-video captioning and rubric scoring, Astra performed better. The findings make task-specific performance important alongside headline cost.

“Research or evaluation purposes only.”

— Nvidia, describing the Flora license

Cost Claims and License Limits

The $0.72-per-hour estimate depends on assumed GPU rental prices and modeled throughput. Nvidia’s paper gives a range, excludes storage and overhead, and says the production GPU-hour denominator is unverified. The comparison also uses only about 75 minutes of video, selected and evaluated by Nvidia. It is not clear how the estimate would change across different footage, workloads, infrastructure or operating conditions.

Benchmark outcomes are mixed, and most tests were conducted by Nvidia. The supplied material does not establish how Flora performs across a broad range of independently selected robot datasets. The license language also means commercial rights should not be assumed: the models are limited to research or evaluation, and derivatives inherit that limit. The terms of any separate commercial permission are not specified in the source material.

Nvidia’s stated capacity of up to 200,000 video hours per day is likewise a company claim about a particular 512 GB300 GPU configuration. The materials do not establish how that capacity translates to typical users or production systems.

Independent Tests and Commercial Terms

The next useful evidence would be independent evaluations on varied robot-video datasets, including comparisons of caption quality, segmentation and reward scoring. Those tests could show whether the performance pattern in Nvidia’s benchmarks holds beyond the episodes used in its paper.

For organizations considering deployment, the immediate next step is to check the model cards and license terms before using Flora beyond research or evaluation. Commercial users will need to establish whether Nvidia offers separate permission, while also calculating their own hardware, storage and operations costs. Until those questions are answered, the reported ratio is best treated as a modeled comparison, not a quoted production price.

Key Questions

What is Nvidia Flora?

Flora is a suite of three models for identifying action intervals in video, writing captions for them and scoring robot actions or outcomes.

Is Flora really 125 times cheaper than GPT-6 Astra?

Nvidia reports a ratio of about 122.5 to 1, rounded to approximately 125×, for captioning on its test set. Flora’s figure is a modeled GPU cost; Astra’s is based on measured API usage. The comparison does not establish a general production-cost ratio.

Can a company use Flora commercially?

The supplied model terms state that Flora is for research or evaluation purposes only, and that derivatives inherit the restriction. The source does not specify whether separate commercial permission is available.

Does Flora outperform GPT-6 Astra on every benchmark?

No. Nvidia reports Flora ahead on some measures, including human-video segmentation and partial-success scoring, while Astra leads on robot-video caption accuracy and rubric scoring.

What does the 125× comparison cover?

It covers the estimated cost of video captioning on 50 episodes totaling about 75 minutes. It does not include Flora’s reward model or establish total operating costs for a production service.

Source: ThorstenMeyerAI.com