🔍 Read the full analysis: What The License Means For Nvidia Flora’s Robot Video Labeling Savings on ThorstenMeyerAI.com
Get the latest gadgets 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 released Flora, three models for segmenting, captioning and scoring robot video. Nvidia estimates captioning costs $0.72 per video hour compared with $88.49 for GPT-6 Astra, but the comparison pairs a modeled GPU budget with measured API usage, and the models are licensed for research or evaluation—not commercial use.
Nvidia’s GEAR robotics lab has released Flora, three models designed to segment, describe and score robot video, and estimates that one captioning task can cost $0.72 per video hour compared with $88.49 using GPT-6 Astra. The headline saving is not a like-for-like measured bill: Nvidia’s Flora figure is a modeled GPU cost, and the models’ license permits research or evaluation only, limiting their direct use in commercial production.
Flora consists 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 grades steps or entire episodes. The models are built on Alibaba’s Qwen3.5 models. Nvidia says the system 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 cost comparison in the paper uses 50 episodes totaling 4,483 seconds, or about 75 minutes of video. GPT-6 Astra’s cost was reported as $88.49 per video hour based on public API rates and usage, with 106 of 7,062 attempts missing usage reports. Flora’s $0.72 figure was calculated using eight GB300 GPUs, an assumed price of $10 per GPU-hour and modeled throughput. Nvidia gives a range of $0.51 to $1.26 per hour, excluding storage and overhead, and says the production GPU-hour denominator has not been verified.
Dividing the two headline figures gives a ratio of about 122.5 to one, which Nvidia rounds to approximately 125×. That comparison concerns captioning, not all three Flora models. On the same episodes, Flora scored 57.1% and GPT-6 Astra 53.7%, according to Nvidia’s results. The paper’s wider benchmarks are mixed: Flora performs better on some human-video, action-boundary and partial-success tests, while Astra leads on robot-video caption accuracy and rubric scoring.
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 Labeling Costs, With Limits
Robot-learning systems depend on video labels that show what action happened, when it happened and whether it succeeded. Producing those labels at scale can consume time and money, so cheaper annotation could let robotics teams process more recordings or direct resources toward data collection and model development. Flora combines segmentation, captions and reward scoring in one set of models, addressing several parts of that workflow.
But the estimate does not establish that a company can run Flora in production for 72 cents per video hour. Actual costs will depend on hardware access, utilization, staffing, storage and other operating expenses; the stated estimate omits storage and overhead. The noncommercial license is a separate practical barrier: organizations cannot treat the released models as a free commercial labeling service under the stated terms. For businesses, Flora may be useful for research, evaluation or as a reference for building other tools, but commercial deployment would require a different permission or licensing arrangement.
For robotics researchers and users, lower labeling costs could ease one constraint on training robots for household and service tasks. It does not resolve the other major challenges highlighted by the source, including collecting useful robot data, hardware expense and safety. Nor does a benchmark score alone establish how well the system will label different tasks or video sources outside the tested set.
NVIDIA GPU workstation for AI model inference
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Flora Labels Robot Video
Robot models can learn from recordings of tasks such as placing an object, passing a component between hands or folding fabric. The Flora paper describes three kinds of labels used in that work: timestamped captions to identify and describe actions, process rewards to grade individual steps, and outcome rewards to score a whole attempt. Those outputs can help researchers organize training data and identify failures.
The three Flora models divide those functions. TempSeg finds the start and end of actions; Captioner uses video context around each interval to write an English instruction; Rewarder can score an episode from one to five, label half-second steps as normal, suboptimal or failure, or apply a user-provided rubric. The model materials describe inference code in the Hugging Face repositories. The source says testing used an Nvidia H100 on Linux and estimates that each 27-billion-parameter model requires roughly 55 GB for weights at 16-bit precision, putting it beyond typical gaming-PC capacity.
The paper’s cost test was conducted on Nvidia’s DenseCaptionBench, with 20 human-hand episodes, 15 involving two-finger grippers and 15 involving dexterous robot hands. The source reports that Nvidia ran most tests itself. The headline price comparison therefore reflects a defined, vendor-selected test set and should be read alongside the paper’s broader results, rather than as a universal price or quality guarantee.
“Research or evaluation purposes only.”
— Nvidia’s OneWay Noncommercial License
As an affiliate, we earn on qualifying purchases.
Cost and Quality Outside the Test
The $0.72 figure is a modeled estimate, not an observed production bill. The source does not provide a verified production GPU-hour denominator, and the estimate excludes storage and overhead. The comparison also sets a measured API charge against a modeled GPU budget, so it does not establish what either system would cost across a company’s full workflow or at different scales.
Quality results are not uniform. Flora’s reported score on the 50 episodes is higher than Astra’s, but other benchmark results favor Astra on robot videos, caption accuracy and rubric scoring. The source does not establish how representative Nvidia’s test episodes are of all robot tasks, or how results change across camera setups, domains and deployment conditions. The full commercial implications of the license may also depend on a user’s intended activity and the applicable terms; no commercial-use exception is reported in the source.
robotics development kit with camera and sensors
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
License Terms and Production Costs
The next practical questions for organizations are whether Nvidia offers a separate commercial license, whether future cost figures include measured production throughput, and how Flora performs on independent or broader evaluations. Teams considering the models will need to review the OneWay Noncommercial License and model cards before use, and account for hardware and operating costs rather than relying on the headline estimate alone.
The source reports no announced date for a commercial license or an independent replication of the cost comparison. Until those details emerge, the 125× figure should be treated as Nvidia’s estimate for a specific captioning test—not as a general guarantee of savings. Further model releases, licensing changes and evaluations could clarify whether Flora can move beyond research and evaluation use.
robotics video annotation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Nvidia Flora?
Flora is a set of three models for segmenting robot video into actions, writing captions for those actions and scoring robot steps or episodes.
How did Nvidia calculate the 125× cost claim?
Nvidia compared a modeled Flora cost of $0.72 per video hour with $88.49 per hour based on GPT-6 Astra API usage for 50 test episodes. The Flora estimate assumed eight GB300 GPUs at $10 per GPU-hour and modeled throughput; it was not a measured production bill.
Can companies use Flora commercially?
The source says all three models use Nvidia’s OneWay Noncommercial License, which permits research or evaluation purposes only. It does not report a commercial-use license or exception.
Does Flora outperform GPT-6 Astra on every test?
No. Nvidia’s reported result on the 50-episode test favored Flora, but the broader benchmark results are mixed. Flora led on some human-video, action-boundary and partial-success measures; Astra led on robot-video caption accuracy and rubric scoring.
Does the cost estimate include all labeling expenses?
No. The Flora estimate excludes storage and overhead, and the source says its production GPU-hour denominator has not been verified. It also covers captioning rather than the cost of every Flora model or a complete commercial workflow.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
