Nvidia Flora’s Cheaper Robot Video Labeling Comes With Licensing Questions

🔍 Read the full analysis: Nvidia Flora’s Cheaper Robot Video Labeling Comes With Licensing Questions on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get smart everyday buys 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, three models designed to mark actions in video, describe them and score robot performance. Nvidia says one part of the system could label video for an estimated $0.72 per hour, far below the $88.49-per-hour API cost it reports for GPT-6 Astra on the same test set; the comparison uses a modeled Flora GPU cost, and the models are licensed for research or evaluation, not commercial use.

Flora is a set of models built on Alibaba’s Qwen3.5 family. Flora-TempSeg, with 9 billion parameters, identifies when actions begin and end. The 27-billion-parameter Flora-Captioner writes an instruction for each segment, while the 27-billion-parameter Flora-Rewarder grades short steps or an entire episode. Nvidia says Flora served as a data engine for its GR00T-2 robot models; the models were trained on fewer than 10,000 hours of human and robot video, according to the supplied model-card details.

The cost comparison concerns captioning, not all three models. Nvidia tested 50 episodes from its DenseCaptionBench, totaling 4,483 seconds, or about 75 minutes. Its paper reports GPT-6 Astra API usage at $88.49 per video hour, based on public rates dated October 4, 2026. Flora’s $0.72 figure is a modeled GPU budget: eight GB300 GPUs priced at an assumed $10 per GPU-hour, with throughput estimated rather than measured. Nvidia gives a range of $0.51 to $1.26 per video hour, excluding storage and other overhead.

On that test set, the paper reports Flora scored 57.1%, compared with 53.7% for Astra, 38.3% for Gemini 3.8 Flash and 30.4% for Claude Opus 5. Those results do not make Flora the best on every task. The paper reports stronger Flora results on human videos, action boundaries and partial-success judgments, while Astra performs better on robot videos, caption accuracy and rubric scoring. Nvidia conducted most of the testing itself, and the source material notes that some Astra attempts reported no usage, making its cost total a known-usage subtotal.

Lower Labeling Costs, Restricted Use

Video labels help robot developers turn recordings into training examples: they can locate actions, describe what happened and identify where a task went wrong. If the reported cost gap holds under production conditions, cheaper labeling could reduce one expense in building robot datasets. Nvidia says Flora can process up to 200,000 hours of video per day on 512 GB300 GPUs, a company-reported capacity that depends on access to that computing scale.

The headline cost comparison is not a like-for-like production bill. Astra’s figure is based on reported API usage for a small, fixed set of episodes; Flora’s is a modeled estimate using an assumed GPU price and throughput. Nvidia’s paper says the production GPU-hour denominator has not been verified. Its estimate also covers captioning, not the cost of running the reward model or building and maintaining a full labeling pipeline. Storage and overhead are excluded.

The license is a separate constraint for companies evaluating the models. All three are distributed under the NVIDIA OneWay Noncommercial License, which permits research or evaluation and extends that restriction to derivatives, according to the source material. The result is a potential research tool and cost benchmark, rather than a ready-made, freely usable commercial labeling service. Lower annotation expense would not resolve other barriers such as collecting suitable robot data, hardware costs or safety testing.

Three Models for Robot Video

Robot-learning systems use video of people and robots performing tasks, but raw footage is not enough for many training workflows. Developers need time-aligned descriptions, judgments about individual steps and scores for complete attempts. Flora separates those functions across its segmenter, captioner and rewarder, rather than asking one model to provide every label.

The models are available through Hugging Face repositories, with inference code included, according to the supplied account. The reported test setup used an Nvidia H100 running Linux; the 27-billion-parameter models are described as requiring roughly 55 GB for weights at 16-bit precision. That estimate points to data-center-class hardware, not an ordinary consumer computer. Nvidia says the models supported GR00T-2, but the available information does not establish how their benchmark performance translates to other datasets or operating environments.

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

— Nvidia paper, Appendix D

Cost and Benchmark Limits

The $0.72 figure is an estimate, not a measured Flora invoice. Nvidia’s assumed $10 GPU-hour rate and modeled throughput may not match a company’s actual hardware, utilization or operating costs; the paper itself says the production GPU-hour denominator is unverified. The reported scenarios range from $0.51 to $1.26 per video hour and exclude storage and overhead.

The performance comparison also has limits. It uses 50 episodes selected from Nvidia’s benchmark, covers about 75 minutes of footage, and was largely run by Nvidia. The supplied material does not establish whether the same cost and quality results hold for broader production datasets. Results differ by task: Flora leads on some human-video and partial-success measures, while Astra leads on some robot-video, captioning and rubric-scoring measures. The source material does not provide commercial-use permission under Flora’s current license.

Independent Testing and License Terms

The next useful evidence would be independent, larger-scale tests that publish actual GPU utilization, hardware costs, overhead and results across a wider range of robot videos. Those measurements could show whether Flora’s modeled savings persist in a production pipeline and how quality changes across tasks.

For companies, the immediate question is whether Nvidia changes the license or offers a separate commercial arrangement; no such change is established in the supplied material. Until then, organizations can assess Flora for research or evaluation under the stated terms, but should not treat the reported cost as a guaranteed production price or the models as cleared for commercial deployment.

Key Questions

What is Nvidia Flora?

Flora is a set of three video models from Nvidia’s GEAR robotics lab: one segments actions in video, one captions them and one scores robot performance.

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

Nvidia’s paper reports $0.72 per video hour for Flora and $88.49 for Astra on its test comparison, a ratio of about 122.5, rounded to approximately 125. But Flora’s price is a modeled GPU estimate, while Astra’s is based on reported API usage; the comparison is not a verified production-cost measurement.

Can a company use Flora commercially?

The supplied license information says the models are released under the NVIDIA OneWay Noncommercial License for research or evaluation only, with that restriction applying to derivatives. The source material does not establish permission for commercial deployment.

Does Flora outperform GPT-6 Astra on every test?

No. The reported overall scores on Nvidia’s selected test set favor Flora, but results vary by task. Flora performs better on some human-video, action-boundary and partial-success measures; Astra performs better on some robot-video, caption-accuracy and rubric-scoring measures.

Source: ThorstenMeyerAI.com