How Nvidia’s Flora Cuts Robot Video Labeling Costs By 125×—And What Its License Means

🔍 Read the full analysis: How Nvidia’s Flora Cuts Robot Video Labeling Costs By 125×—And What Its License Means on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get movie nights delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Nvidia’s GEAR robotics lab has released Flora, a set of three models that segment, caption and score robot-video recordings, and estimates that its captioning pipeline costs $0.72 per video hour compared with $88.49 for GPT-6 Astra. The claimed advantage is about 122.5×, rounded to 125×, but Nvidia’s Flora figure is a modeled GPU cost rather than a measured production bill, and the models’ license restricts use to research or evaluation.

Flora includes a 9-billion-parameter temporal segmenter that identifies when actions begin and end, a 27-billion-parameter captioner that describes each segment, and a 27-billion-parameter reward model that scores an episode or grades steps. The models are built on Alibaba’s Qwen3.5 family. Nvidia says Flora was used as a data engine for its GR00T-2 robotics model and was trained on fewer than 10,000 hours of human and robot video.

The cost comparison in the paper uses 50 episodes totaling 4,483 seconds, or about 75 minutes, from Nvidia’s DenseCaptionBench. GPT-6 Astra’s $88.49 per video hour is based on reported public API usage at October 4, 2026 rates; 106 of 7,062 attempts had no reported usage, making the total a known-usage subtotal. Flora’s $0.72 figure is calculated using eight GB300 GPUs, assumed to cost $10 per GPU-hour, and modeled throughput. Nvidia lists scenarios from $0.51 to $1.26 per hour, excluding storage and overhead.

The paper reports an overall score of 57.1% for Flora on the test episodes, compared with 53.7% for Astra, 38.3% for Gemini 3.8 Flash and 30.4% for Claude Opus 5. Results vary by task: Flora did better on the human-video subset, action segmentation and some cases involving partial success, while Astra scored higher on robot videos, caption accuracy and rubric scoring. The 125× comparison covers captioning, not Flora’s reward model.

Lower Costs, With Limits

Video labeling is a costly step in building robot-training datasets. Recordings have to be divided into actions, described in language and, for some training methods, scored for success or failure. If automated labeling is cheaper at scale, robotics teams could process more recorded data within the same budget and make it easier to identify useful training examples.

But the headline ratio does not establish that companies can run Flora in production at that price. The comparison places a modeled GPU budget against API usage measured for a different system, using a limited test set. Nvidia’s paper says the production GPU-hour denominator has not been verified. Hardware, engineering, storage and operational costs can also affect an actual deployment. The models’ noncommercial license further means businesses cannot treat the release as a free production service.

For researchers, Flora offers downloadable models and a comparison point for video-labeling systems. For robotics companies, the benchmark may indicate where lower-cost labeling could help, but it does not remove other constraints, including collecting useful robot data, building safe systems and paying for suitable hardware.

Three Models for Robot Video

Robot-learning systems use recordings of tasks such as placing an object or folding fabric. Labels can specify what happened and when, grade individual steps, or score the outcome of an entire attempt. Those annotations help teams organize training data and examine where a robot’s behavior went wrong.

Flora divides those jobs across its three models. The segmenter returns time intervals; the captioner writes an English instruction for each interval and can use multiple camera views; the rewarder can provide an episode score from 1 to 5, grade steps as normal, suboptimal or failure, or evaluate a user-provided rubric. The project’s model cards list inference code in the Hugging Face repositories. The source report says testing used an Nvidia H100 on Linux and estimates that each 27B model needs roughly 55 GB for weights in 16-bit precision, pointing to data-center-class hardware rather than an ordinary gaming PC.

The licensing terms are a separate part of the release. The three models are distributed under the NVIDIA OneWay Noncommercial License, which permits research or evaluation and says derivatives inherit the restriction. ThorstenMeyerAI.com reports that Nvidia may use the models commercially; the release therefore does not grant outside companies the same commercial rights.

“research or evaluation purposes only”

— NVIDIA OneWay Noncommercial License, as quoted by ThorstenMeyerAI.com

Production Cost and Scope

The reported ratio is not a like-for-like measurement of two production deployments. Flora’s cost is modeled using an assumed GPU rate and throughput; the paper says the production GPU-hour denominator remains unverified. Astra’s figure comes from measured API usage on the selected episodes, with some attempts lacking usage reports. The cost scenarios also exclude storage and overhead.

The benchmark was run on 50 episodes selected from Nvidia’s own DenseCaptionBench, and Nvidia conducted most of the tests, according to the source report. The results show Flora leading on some tasks but not others. It is not clear from the supplied material how the systems would compare across larger or independently selected datasets, or what costs and performance a company would see with its own hardware and workflow. The estimate does not include reward-model costs, and the source material does not provide a verified commercial deployment cost.

Independent Tests and License Terms

The next useful evidence would be independent evaluations on larger and varied robot-video datasets, alongside measured costs from real deployments that account for hardware utilization, storage and operational overhead. Those results could show whether Flora’s modeled savings hold outside the test setup and how its mixed benchmark performance affects practical use.

Organizations considering the models will also need to review the license before deployment. Under the terms described in the release, Flora is available for research or evaluation, not commercial use by outside companies. No commercial licensing arrangement, verified production cost or further release milestone is specified in the source material.

Key Questions

What is Nvidia Flora?

Flora is a group of three models for processing robot video: one segments actions in time, one captions them, and one scores robot performance.

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

Nvidia’s paper reports a ratio of 122.5×, rounded to 125×, using $0.72 per hour for Flora and $88.49 for Astra. Flora’s figure is modeled from assumed GPU costs and throughput, while Astra’s is based on API usage on 50 episodes; it is not a verified comparison of production bills.

Can companies use Flora commercially?

The models are distributed under the NVIDIA OneWay Noncommercial License, described as permitting research or evaluation only. The supplied report says derivatives inherit that limit, so outside companies should not assume the release permits commercial deployment.

Did Flora outperform GPT-6 Astra on every test?

No. Flora scored higher overall on the reported test set and led on some human-video and action-segmentation results. Astra did better on robot videos, caption accuracy and rubric scoring, according to the source report.

What does the 125× estimate leave out?

The Flora estimate excludes storage and overhead, and the comparison covers captioning rather than the reward model. The reported figure is based on modeled GPU use, not a verified production deployment.

Source: ThorstenMeyerAI.com