Robot Video Labeling Costs Could Fall 125× With Nvidia Flora—License Included

🔍 Read the full analysis: Robot Video Labeling Costs Could Fall 125× With Nvidia Flora—License Included 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, three models designed to segment, caption and score robot videos, alongside a paper estimating that its captioning costs could be about 122.5 times lower than GPT-6 Astra’s on a selected test set. The comparison is not a like-for-like bill: Nvidia’s Astra figure comes from measured API usage, while Flora’s $0.72 per video hour is a modeled GPU estimate, and the models are licensed for noncommercial research or evaluation.

Flora consists of three models built on Alibaba’s Qwen3.5 family. 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, using video context from before and after the action. Flora-Rewarder can score an episode, grade short steps as normal, suboptimal or failures, or assess performance against a user-provided rubric.

Nvidia’s cost comparison used 50 episodes totaling 4,483 seconds, or about 75 minutes, from its DenseCaptionBench: 20 human-hand videos, 15 videos of two-finger grippers and 15 of dexterous robot hands. The paper reports GPT-6 Astra at $88.49 per video hour, based on API usage at public rates, and Flora at $0.72. Nvidia calculated Flora’s figure using eight GB300 GPUs, an assumed price of $10 per GPU-hour and modeled throughput; its scenarios ranged from $0.51 to $1.26 per hour.

Dividing the reported Astra cost by the modeled Flora cost gives about 122.5 times, rounded in Nvidia’s headline figure to approximately 125×. The paper says 106 of 7,062 Astra attempts had no reported usage, so the API total is a known-usage subtotal. Nvidia also states that the assumed production GPU-hour denominator has not been verified. The comparison concerns captioning costs; it does not include Flora-Rewarder.

Lower Costs, With Limits

Video labeling is a data-preparation expense for robotics teams: timestamped descriptions help organize training footage, while step-level and episode-level scores can help identify failures or evaluate results. If inference costs fall substantially, teams could process more recordings within the same budget. Nvidia says Flora served as a data engine for its GR00T-2 robot models, showing how the company says it has used the tools in its own development pipeline.

The price comparison should not be read as proof that a company can label production data for 72 cents per hour. Flora’s cost is a projection based on specified GPU assumptions, and the test covers a limited set chosen by Nvidia. The paper’s reported quality results are also mixed: Flora scores higher than Astra on the combined DenseCaptionBench figure in the supplied comparison, but Astra leads on some robot-video and caption-accuracy measures. Hardware, storage and operational overhead also affect real deployment costs.

For businesses, the noncommercial license is a direct constraint. The model cards specify research or evaluation use, and derivatives inherit that restriction. The release may provide researchers with tools to test and compare video-labeling methods, but it does not, on the stated terms, offer companies a free commercial labeling service. Lower labeling costs would address one expense in robotics; data collection, hardware and safety remain separate challenges.

Three Models for Robot Video

Robot-learning datasets include recordings of tasks such as moving objects, manipulating tools or folding fabric. Turning those recordings into useful training data can require labels that say what happened and when, as well as judgments about whether each step or the overall attempt succeeded. Flora divides those functions across its segmenter, captioner and rewarder rather than relying on a single output.

According to Nvidia’s paper and model cards, Flora was trained on under 10,000 hours of human and robot video. The models are available through Hugging Face repositories with inference code. Nvidia lists an H100 running Linux as test hardware. The source material estimates that each 27-billion-parameter model requires roughly 55 GB for weights at 16-bit precision, a scale more suited to data-center hardware than a typical personal computer.

Nvidia’s reported benchmark results show different strengths across tasks. On the public WGO-Bench segmentation test, Flora scored 50.3 versus Astra’s 46.0, and it scored 37.4 versus 29.7 on the human-video subset of DenseCaptionBench. On the public RoboRewardBench, the paper reports Flora’s error as 27.8% lower than Astra’s. These results do not establish that Flora leads on every task: the supplied source says Astra performs better on robot videos, caption accuracy and rubric scoring.

“The production GPU-hour denominator has not been verified.”

— Nvidia, in the Flora paper

Cost and Performance Caveats

The 125× comparison is an estimate, not an independently verified production-cost result. Flora’s GPU price assumes $10 per GPU-hour and modeled throughput; the cited range is $0.51 to $1.26 per video hour, excluding storage and overhead. Astra’s figure comes from API usage on the same 50 episodes, but 106 attempts lacked usage reports. The small, vendor-selected test set also cannot establish how costs or quality would compare across other datasets and workloads.

Quality varies by task. The source reports Flora ahead on human videos, action boundaries and partial-success scoring, while Astra leads in some robot-video, caption-accuracy and rubric-scoring evaluations. It does not provide enough detail to determine how those differences would affect a particular robotics company’s workflow. The license further leaves commercial users without permission to deploy the models commercially under the terms described.

Testing Beyond Nvidia’s Benchmark

The next useful evidence would be independent evaluations on broader robot-video datasets and measured cost reports from production deployments. Those could show whether Flora’s modeled savings hold when teams account for hardware utilization, storage, engineering work and the mix of tasks they need labeled. Until then, the published comparison is a reference point rather than a guaranteed operating cost.

Researchers and other permitted users can test the released models for research or evaluation. Companies considering commercial use will need to review the license and seek appropriate permission or alternatives. Nvidia’s paper and model cards provide the current benchmark and deployment details; no separate commercial-use authorization or follow-up production measurement is described in the supplied material.

Key Questions

What does Nvidia Flora do?

Flora is a set of three video models: one finds action boundaries, one writes captions for video segments, and one scores steps or complete episodes.

Is the 125× cost reduction a measured result?

No. Nvidia’s $0.72 per video hour for Flora is modeled using assumed GPU pricing and throughput. The GPT-6 Astra figure of $88.49 per hour is based on reported API usage for the test episodes, with some attempts lacking usage data.

Can a company use Flora commercially?

The supplied model cards specify the NVIDIA OneWay Noncommercial License, limited to research or evaluation, with derivatives inheriting the restriction. Commercial use is not permitted by the described license.

Does Flora outperform GPT-6 Astra on every benchmark?

No. The reported results are mixed by task. Flora leads on some human-video, segmentation and partial-success measures, while Astra leads on some robot-video, caption-accuracy and rubric-scoring evaluations.

Source: ThorstenMeyerAI.com