Read the full analysis: Robot Video Labeling For Less: Nvidia’s Flora And Its 125× Cost Claim on ThorstenMeyerAI.com
TL;DR
Nvidia’s GEAR lab released Flora, three models for segmenting, captioning and scoring robot videos, and reports a labeling cost of $0.72 per video hour. Its approximately 125× savings figure compares a modeled GPU cost with a measured GPT-6 Astra API bill; benchmark results vary, and the models are licensed for noncommercial research and evaluation.
Recommended by digitechbytes.com · AI-assisted guides
Shopping for emerging consumer tech explained? Start with the guides we keep up to date:
Updated October 202615 Best Robot Vacuum Cleaners for Pet Hair in 2027See the top picks →Updated October 202615 Best Robot Vacuum Cleaners for Easier Floor Care in 2027See the top picks →Updated October 202614 Best Smartphone Gimbals for Halloween Videos in 2027See the top picks →
Nvidia’s GEAR robotics lab has released Flora, a set of three models that segment, describe and score robot videos, alongside a claim that labeling costs about $0.72 per video hour—roughly 125 times less than using GPT-6 Astra. The comparison is Nvidia’s estimate, not a direct comparison of two measured production bills, and Flora’s license limits use to research or evaluation.
Flora comprises three models built on Alibaba’s Qwen3.5 family. Flora-TempSeg, with 9 billion parameters, identifies where actions begin and end in a video. The 27-billion-parameter Flora-Captioner writes an instruction describing each segment. The similarly sized Flora-Rewarder grades a full episode from 1 to 5, evaluates steps as normal, suboptimal or failed, or scores performance against a user-provided rubric.
Nvidia’s paper says Flora was used as a data engine for its GR00T-2 robot models and was trained on fewer than 10,000 hours of human and robot video. The lab reports that its system can process up to 200,000 hours of video per day using 512 GB300 GPUs. That is a Nvidia-reported capacity figure; the release does not establish that this throughput has been independently verified in production.
The cost comparison uses 50 episodes from Nvidia’s DenseCaptionBench, totaling 4,483 seconds, or about 75 minutes. Nvidia reports $88.49 per video hour for GPT-6 Astra API use and models Flora at $0.72 per hour, based on eight GB300 GPUs priced at an assumed $10 per GPU-hour. Dividing those figures gives 122.5, rounded to approximately 125×. Nvidia says its Flora estimate ranges from $0.51 to $1.26 per hour and excludes storage and overhead; the paper states that the production GPU-hour denominator has not been verified. The comparison covers captioning, not Flora’s reward model.
Nvidia GEAR · Oct 2026
Infographic 01 · What Flora does
Three models that turn robot video into training labels.
Flora cuts a recording into actions, describes each one and grades how well it went. The 200,000-hour figure is Nvidia’s own production claim.
Flora-TempSeg
9B · adapter on Qwen3.5-9B
Job: Finds where each action starts and endsIn: Video windows (one camera, or main + two wrist cameras)Out: Time intervals as JSON
Flora-Captioner
27B · fine-tuned Qwen3.5-27B
Job: Describes each action, using the video before and afterIn: One interval plus contextOut: One English instruction, e.g. “Place the blue mug on the tray with your right hand.”
Flora-Rewarder
27B · fine-tuned Qwen3.5-27B
Job: Judges how well the robot didIn: An episode or a 16-second window, plus the taskOut: Score 1–5 per episode · normal / suboptimal / failure per 0.5 s · rubric scores
Three labels, three uses
Why robot models need each one, per the paper
What it is not
model cards
Not a robot controller: outputs are annotations onlyNot meant to be the sole safety monitor for a physical systemNot for commercial or production use (noncommercial license)Hardware: tested on Nvidia H100, Linux; ~55 GB of weights per 27B model (estimate)
Sources: Flora paper, Nvidia (20 Oct 2026); Hugging Face model cards (15 Oct 2026); Nvidia GEAR project page. Weight-size estimate: Thorsten Meyer (27B parameters × 2 bytes).
Powered by Thorsten Meyer AI
Lower Labeling Costs, With Limits
Video labels help robotics teams identify actions, train models and locate failures in long recordings. Reducing the cost of producing them could make it easier to work with larger training datasets, particularly when human annotation or repeated use of a general-purpose model is costly. For developers, the release also offers a set of models aimed at different stages of the labeling process rather than a captioner alone.
But the cost headline does not establish what a company would pay in a live deployment. Flora’s figure is modeled, while Astra’s is based on reported API usage; the test was short and selected by Nvidia. Quality also varies by task. On the same episodes, Nvidia reports a combined score of 57.1% for Flora and 53.7% for Astra, but Astra performed better on robot videos, caption accuracy and rubric scoring. The results suggest a possible lower-cost tool, not a general quality lead.
The license is another practical constraint: the models are available for noncommercial research and evaluation, not as an unrestricted production pipeline. Companies may use the release to assess approaches or benchmark systems, but the stated terms do not grant commercial use of Flora itself.
Three Models for Video Labels
Robot-learning datasets include recordings of machines grasping objects, manipulating tools or failing at a task, as well as videos of people performing similar actions. To use that material, developers may need timestamps for each action, a short description of what happened and a score for how well a step or full attempt went. Those labels can help organize training data and evaluate robot behavior.
Flora separates those jobs across its three models. The captioner can use a main camera and, where available, wrist-camera views. Nvidia’s model cards list inference code in the Hugging Face repositories and describe testing on an Nvidia H100 running Linux. The supplied source estimates that each 27-billion-parameter model needs roughly 55 GB for weights in 16-bit precision, putting local use beyond the capacity of many consumer computers.
The released models use the NVIDIA OneWay Noncommercial License. Its stated scope is research or evaluation, with derivatives subject to the same restriction. Nvidia’s own commercial use is not barred by the license, according to the supplied source. The paper and model cards describe the release; the cost and benchmark findings remain Nvidia’s reported results.
“Research or evaluation purposes only.”
— NVIDIA OneWay Noncommercial License
Cost and Quality Still Need Testing
It is not clear how Flora’s modeled hourly cost would compare with a measured production bill after including infrastructure, storage, engineering and operational overhead. Nvidia’s test covered 50 episodes and about 75 minutes, and its paper says the production GPU-hour denominator has not been verified. The estimate also depends on the assumed $10 hourly price for each GB300 GPU.
The benchmark results do not show Flora leading on every task. Nvidia reports wins for Flora on human videos, action boundaries and partial-success evaluation, while GPT-6 Astra scored better on robot-video captioning and rubric-based scoring. Most testing was conducted by Nvidia, and the supplied material does not describe an independent replication. It is also unclear how results would change across other datasets, camera setups and longer videos.
The source gives no independently audited confirmation of the reported daily throughput. Nor does the cost comparison include the reward model, so it should not be read as the total cost of every labeling function Flora provides. Commercial licensing options beyond the stated noncommercial terms are not described.
Independent Tests and License Terms
The immediate next step is for robotics teams and researchers to test the models against their own videos and annotation standards, while accounting for hardware and operating costs. Independent evaluations could clarify whether Flora’s reported quality and speed hold outside Nvidia’s benchmark and whether the modeled cost is achievable under production workloads.
Organizations considering deployment will also need to review the license before using the models or derivatives in commercial products. Nvidia’s release provides models, cards and a paper, but the supplied information does not announce a commercial license or a date for further validation. Until those details emerge, the 125× figure remains a useful estimate to examine—not a verified production saving.
Key Questions
What does Nvidia’s Flora do?
Flora is a set of three models: one segments video into actions, one captions those segments, and one scores robot performance by step, episode or a user-provided rubric.
Is the 125× cost saving independently verified?
No independent verification is described in the supplied material. Nvidia compares a modeled Flora cost of $0.72 per video hour with a GPT-6 Astra API figure of $88.49 per hour and says the GPU-hour denominator has not been verified.
Did Flora outperform GPT-6 Astra?
Not across every task. Nvidia reports a higher combined score for Flora on its 50-episode test, but says Astra did better on robot videos, caption accuracy and rubric scoring. Flora performed better on some human-video, segmentation and partial-success measures.
Can companies use Flora commercially?
The stated NVIDIA OneWay Noncommercial License permits research or evaluation only, and derivatives inherit that limit. The supplied material does not identify a commercial-use license for other organizations.
Source: ThorstenMeyerAI.com