🔍 Read the full analysis: Can Nvidia’s Flora Deliver 125× Cheaper Robot Video Labeling? on ThorstenMeyerAI.com
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
TL;DR
Nvidia’s GEAR lab has released Flora, a three-model system for segmenting, describing and scoring robot video. Nvidia estimates captioning costs at $0.72 per video hour versus $88.49 for GPT-6 Astra, but the ratio compares modeled GPU costs with measured API usage on a small, Nvidia-selected test set; Flora’s license also bars commercial use.
Nvidia’s GEAR robotics lab has released Flora, a set of three models that segment, caption and score robot videos, and estimates that its captioning costs $0.72 per video hour compared with $88.49 for GPT-6 Astra on a test set. The headline saving is a modeled estimate—not a direct comparison of two measured production bills—and the models are released under a noncommercial license.
Flora is made up of Flora-TempSeg, a 9-billion-parameter model that identifies action intervals; Flora-Captioner, a 27-billion-parameter model that writes an instruction for each interval; and Flora-Rewarder, a 27-billion-parameter model that scores an episode or grades steps. The models are built on Alibaba’s Qwen3.5 family, according to the paper and model cards described by ThorstenMeyerAI.com.
The cost estimate applies to captioning, not all three Flora functions. Nvidia compared the models on 50 episodes totaling 4,483 seconds, or about 75 minutes, from its DenseCaptionBench. GPT-6 Astra’s $88.49 per video hour reflects reported API usage at public rates; Flora’s $0.72 figure is calculated from modeled throughput on eight GB300 GPUs, priced at an assumed $10 per GPU-hour. Nvidia gives a Flora range of $0.51 to $1.26 per hour, excluding storage and other overhead.
Dividing the two headline figures gives about 122.5 times, which Nvidia rounds to approximately 125 times. The source says 106 of 7,062 Astra attempts reported no usage, making its cost a known-usage subtotal. Nvidia also states that the production GPU-hour denominator has not been verified. The comparison is therefore an estimate based on different cost methods, rather than an independently verified operating-cost result.
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 ends
- In: 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 after
- In: One interval plus context
- Out: 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 did
- In: An episode or a 16-second window, plus the task
- Out: 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
| Label | Used for | Made by |
|---|---|---|
| Timestamped captions | Filtering and rebalancing pretraining data | TempSeg + Captioner |
| Process rewards | Reinforcement learning after pretraining: which step went wrong | Rewarder |
| Outcome rewards | Monitoring robots after deployment | Rewarder |
What it is not
model cards
- Not a robot controller: outputs are annotations only
- Not meant to be the sole safety monitor for a physical system
- Not 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-Cost Labels, With Limits
Robot-learning systems need video labels to connect what a robot sees with the actions it takes and the results it gets. Timestamped captions can help teams search and balance training data, while step-level failure scores can help identify where a task went wrong. If Flora can produce those labels at lower cost and useful quality, it could reduce one expense in building datasets for robot models.
The evidence in the cited benchmark is mixed rather than a clear overall win. The source reports that Flora scored 57.1% on Nvidia’s 50-episode test, compared with 53.7% for GPT-6 Astra. It also reports Flora performed better on human-hand videos, action boundaries and partial-success cases. Astra scored higher on robot videos, caption accuracy and rubric scoring. These results come from tests Nvidia largely conducted itself, and they do not establish how the models perform across different datasets or in commercial deployments.
Access terms also shape the practical impact. Flora’s models use the NVIDIA OneWay Noncommercial License, described as permitting research or evaluation only, with derivatives subject to the same restriction. The source says Nvidia may use the models commercially. For companies, the release can provide a research comparison and technical reference, but the stated license does not make Flora a ready-to-use, free commercial labeling service.
Popular right now
Vitamix 5200 Professional-Grade Blender, 64 oz. Container, Black
TCL 85-Inch QM6K Series QLED Mini LED 4K HDR Smart Google TV
Ninja Foodi Air Fryer 10 Qt DualZone 6-in-1 Black DZ302
As an affiliate, we earn on qualifying purchases.
Three Models for Robot Video
Robot models are trained using recordings that can include a gripper placing an object, two robot hands passing a part, or a person carrying out a task. Flora is designed to turn those recordings into several kinds of supervision: action boundaries, descriptions tied to time intervals and scores for steps or whole attempts. The three-model design separates those functions rather than asking one model to produce every label.
According to the source’s account of the model cards, Flora was used as a data engine for Nvidia’s GR00T-2 and was trained on fewer than 10,000 hours of human and robot video. The models are available through Hugging Face repositories, with inference code included there. Testing described in the source used an Nvidia H100 running Linux; the 27-billion-parameter models require substantial memory, which puts local use beyond the typical consumer computer.
The cost test included 20 human-hand episodes, 15 with two-finger grippers and 15 with dexterous robot hands. The source reports other benchmark results too: Flora scored 50.3 on WGO-Bench segmentation versus Astra’s 46.0, and Flora’s error on RoboRewardBench was 27.8% lower. Those separate metrics offer useful comparisons, but they do not remove the limits of the selected test set or settle performance across real-world tasks.
“The production GPU-hour denominator has not been verified.”
— Nvidia, as quoted in the Flora paper
Cost and Quality Still Unsettled
The headline ratio depends on a modeled Flora GPU cost and measured Astra API usage, so it is not a like-for-like production comparison. The assumed GPU rental rate, modeled throughput and exclusion of storage and overhead affect the estimate. Nvidia’s paper itself says the production GPU-hour denominator has not been verified; the cost of running Flora at sustained commercial scale remains unknown.
The benchmark covers only 50 episodes, about 75 minutes of footage, and the source says Nvidia selected and scored the test. Results may differ with other video, robot types, evaluation methods or production workloads. The provided material does not establish independent replication, a broad real-world quality comparison, or the cost of labeling with Flora’s reward model.
The commercial implications are also limited by the license. The source describes it as noncommercial, but the supplied material does not detail any separate commercial licensing route or its terms. It is also unclear whether the reported cost and quality results will hold when the models process much larger datasets under varied operating conditions.
Independent Tests and License Terms
The next useful evidence would be independent testing on larger and more varied robot-video datasets, alongside a measured accounting of GPU time, hardware utilization, storage and operating overhead. That would help distinguish a promising cost model from the cost of a sustained labeling pipeline.
Organizations considering Flora will also need to review the license before using the models beyond research or evaluation. The supplied source does not announce a commercial release, a revised license or a planned independent benchmark. Until those details are available, the 125-times figure should be read as Nvidia’s estimate for the stated captioning test—not a guarantee of savings for every dataset or deployment.
Key Questions
What is Nvidia Flora?
Flora is a three-model system for robot video: one model segments action intervals, one captions them, and one scores steps or full episodes.
Is Flora proven to be 125 times cheaper?
No. The approximately 122.5-times cost ratio comes from comparing Nvidia’s modeled Flora GPU cost with measured GPT-6 Astra API usage on 50 episodes. Nvidia says the production GPU-hour denominator has not been verified.
Does the cost estimate cover all of Flora?
No. The $0.72-per-hour estimate covers captioning. The provided source says the Flora reward model is not included in the 125-times comparison.
Can companies use Flora commercially?
The source describes Flora’s NVIDIA OneWay Noncommercial License as allowing research or evaluation only, with derivatives inheriting the restriction. It does not provide terms for a separate commercial license.
Did Flora outperform GPT-6 Astra on every test?
No. Flora scored higher on some measures, including human videos and partial-success cases, while Astra scored higher on robot videos, caption accuracy and rubric scoring, according to the source’s account of the benchmark results.
Source: ThorstenMeyerAI.com
