The Trade-Off In Nvidia Flora’s Cheaper Robot Video Labeling

🔍 Read the full analysis: The Trade-Off In Nvidia Flora’s Cheaper Robot Video Labeling 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, a set of three models designed to label video used in robot training, and estimates that its captioning pipeline costs $0.72 per hour of video. That figure is far below the $88.49 reported for GPT-6 Astra on Nvidia’s test, but it is a modeled GPU cost rather than a measured production bill, and Flora’s license permits research and evaluation, not commercial use.

Flora comprises a 9-billion-parameter segmenter that identifies action intervals, a 27-billion-parameter captioner that describes each interval in English, and a 27-billion-parameter rewarder that scores robot attempts. The reward model can grade short steps as normal, suboptimal or failure, score a full episode from 1 to 5, or apply a user-provided rubric, according to the paper and model cards.

Nvidia’s cost comparison uses 50 episodes totaling 4,483 seconds—about 75 minutes—from its DenseCaptionBench. GPT-6 Astra’s $88.49 per video hour is based on reported API usage at public rates from Oct. 4, 2026. Flora’s $0.72 estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and a modeled processing rate. Nvidia gives a range of $0.51 to $1.26 per hour, excluding storage and other overhead.

The paper reports an approximate 122.5-to-1 cost ratio, rounded to about 125 times. It also cautions that the “production GPU-hour denominator has not been verified.” The comparison covers captioning, not all three Flora models, and includes 106 Astra attempts with no reported usage that were left out of the known-usage subtotal.

Lower Labeling Costs, With Limits

Robot systems need more than video footage: they need descriptions of what happened and when, plus signals about whether an action succeeded. Automating those labels could reduce a costly step in preparing training data, making it easier for robotics teams to filter and evaluate larger video collections. Nvidia says Flora served as a data engine for its GR00T-2 robot models.

The reported price gap is meaningful as a cost estimate, but it does not establish what a company would pay in a real deployment. The estimate depends on GPU rental rates and modeled throughput, while the Astra number comes from API billing. Infrastructure, storage, engineering and operating costs are outside the Flora estimate. The headline ratio also says nothing about the cost of collecting robot video or building and testing the robot itself.

Access is another limit for businesses. All three models are distributed under the NVIDIA OneWay Noncommercial License, which allows research or evaluation and applies the same restriction to derivatives. The release can give companies a reference point for labeling methods, but the stated terms do not make Flora a free commercial production service.

How Flora Labels Robot Video

Robot-learning datasets can include recordings from robot cameras as well as videos of people performing tasks. The labels described in Flora’s paper include timestamped captions, which connect instructions to particular portions of a clip; process rewards for individual steps; and outcome scores for an entire attempt. These labels can help researchers organize examples and identify errors.

The three Flora models are based on Alibaba’s Qwen3.5 models. Nvidia says the system was trained on fewer than 10,000 hours of human and robot video. Its paper compares Flora with GPT-6 Astra and other systems on several benchmarks, but the main cost comparison uses Nvidia’s own 50-episode test set. Nvidia also reports processing up to 200,000 hours of video per day on 512 GB300 GPUs; that is the company’s stated capacity, not an independent production measurement.

The results are not a clean sweep. On the reported test set, Flora scored 57.1% and Astra 53.7%, while Gemini 3.8 Flash scored 38.3% and Claude Opus 5 scored 30.4%. Flora performed better on the human-video portion and on some action-boundary and partial-success measures. Astra remained ahead on robot-video caption accuracy and rubric scoring, according to the benchmark results described in the source material.

“Expensive at the scale of modern robot datasets.”

— Nvidia’s description of frontier models

What the Cost Estimate Leaves Out

The $0.72 figure is not a measured production invoice. It depends on an assumed GPU rental price and modeled throughput; the paper says the production GPU-hour denominator has not been verified. The reported range does not include storage or overhead, and the comparison pairs that estimate with measured API usage. How the figures would compare under a specific company’s hardware, workload and operating costs is not established.

Quality results also depend on the task and test set. Nvidia ran most of the evaluations described, and the central comparison uses 50 episodes selected for its DenseCaptionBench. The source material does not establish how well the results generalize to other robot types, camera setups, tasks or longer production datasets. Flora’s advantage on some measures does not mean it outperforms Astra across all labeling work.

The license terms are stated as noncommercial, but the source does not detail what commercial arrangements Nvidia may offer or whether those terms could change. The cost comparison does not include reward-model inference, data collection, safety evaluation or the other expenses involved in deploying robots.

Commercial Use and Broader Testing

Researchers can evaluate the released models under the license terms, while businesses will need to determine whether those terms fit their intended use or seek other permissions. Further independent testing could clarify how Flora performs across datasets and whether its modeled throughput and cost hold under production workloads.

For now, the key next step is comparison on more varied robot videos and transparent accounting that includes the full operating pipeline. Nvidia’s results suggest that automated labeling may lower one cost in robot-data preparation, but they do not settle the total cost or performance trade-offs for commercial systems.

Key Questions

What is Nvidia Flora?

Flora is a set of three video models for robot-data labeling: a segmenter that finds action intervals, a captioner that describes them, and a rewarder that scores steps or full attempts.

Is the $0.72-per-hour figure a measured cost?

No. It is a modeled estimate based on eight GB300 GPUs rented at an assumed $10 per GPU-hour. Nvidia gives a range of $0.51 to $1.26 and says the production GPU-hour denominator has not been verified.

Did Flora outperform GPT-6 Astra?

Results varied by task. Flora scored 57.1% against Astra’s 53.7% on the reported 50-episode test, and Flora led on some human-video and partial-success measures. Astra performed better on robot-video caption accuracy and rubric scoring, according to the reported results.

Can companies use Flora commercially?

The models are released under the NVIDIA OneWay Noncommercial License, which permits research or evaluation. The source material does not describe any separate commercial-use terms.

Does the 125-times comparison include every Flora model?

No. The comparison concerns captioning costs on a 50-episode test set. It does not include the cost of using Flora’s rewarder or the broader expenses of collecting, storing and operating robot datasets.

Source: ThorstenMeyerAI.com