Nvidia Flora Could Make Video Labeling For Robots Far Cheaper—But Check The Terms

🔍 Read the full analysis: Nvidia Flora Could Make Video Labeling For Robots Far Cheaper—But Check The Terms on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Nvidia’s GEAR robotics lab has released Flora, a set of three models designed to segment, describe and score video of robot activity. Nvidia estimates its captioning pipeline could process video for $0.72 per hour, far below the $88.49 per hour it reports for GPT-6 Astra on the same test episodes; the Flora figure is a modeled GPU cost, not a measured production bill, and its license bars commercial use.

Flora includes a 9-billion-parameter temporal segmenter that identifies where actions begin and end, a 27-billion-parameter captioner that writes instructions for each interval, and a 27-billion-parameter reward model that grades steps or whole robot attempts. Nvidia says the models were built on Alibaba’s Qwen3.5 models and that Flora served as a data engine for its GR00T-2 robot models. The model cards say the training used fewer than 10,000 hours of human and robot video.

The cost comparison in the Flora paper uses 50 episodes totaling 4,483 seconds, or about 75 minutes, from Nvidia’s DenseCaptionBench. Nvidia reports GPT-6 Astra API costs of $88.49 per video hour, based on public rates from October 4, 2026. Flora’s $0.72 estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput. The paper gives a range of $0.51 to $1.26 per hour and says storage and overhead are excluded. Its stated ratio of about 125 times is the $88.49 API figure divided by the $0.72 estimate.

The comparison covers captioning costs, not every Flora model or the full expense of building a robot-data pipeline. Nvidia’s paper says the production GPU-hour denominator has not been verified. It also reports that 106 of 7,062 API attempts had no usage data, making the Astra cost a known-usage subtotal. The quality results on the test episodes are mixed: Flora scored 57.1%, Astra 53.7%, Gemini 3.8 Flash 38.3% and Claude Opus 5 30.4%, according to Nvidia’s evaluation.

Lower Labeling Costs, With Caveats

Labeling video is a costly step in training robot systems: teams need to identify actions, attach descriptions to the relevant moments, and judge whether a robot’s attempt succeeded. If Flora can deliver usable labels at a lower cost, it could reduce one obstacle to turning large collections of video into training data. That could help researchers and developers test robot-learning systems across more examples.

But the headline saving is not a direct comparison of two measured production bills. It sets a modeled GPU budget against an API charge, using a small benchmark selected and evaluated by Nvidia. The result is useful as an estimate, but it does not establish what a company would pay at scale after hardware, storage, engineering and operating costs. The quality findings also do not show Flora leading on every task: Nvidia reports advantages on human videos, action boundaries and partial successes, while Astra performs better on robot videos, caption accuracy and rubric scoring.

The license is another practical limit. The models are released under the NVIDIA OneWay Noncommercial License, which permits research or evaluation use, not ordinary commercial deployment. That makes Flora available as a research tool and comparison point, but not automatically as a free production service for robotics companies.

Three Models for Robot Video

Robot-learning systems use recordings of tasks such as placing an object or folding fabric. To make those recordings useful for training, developers need labels that describe what happened and when. Flora’s segmenter identifies action intervals; its captioner supplies an instruction for each interval; and its rewarder assigns scores, including half-second grades of normal, suboptimal or failed behavior. The reward model can also assess an episode against a user-provided rubric.

Nvidia says Flora was developed by its GEAR lab, associated with Yuke Zhu and Linxi “Jim” Fan. The models and inference code are available through Hugging Face repositories, according to the source material. Testing used an Nvidia H100 on Linux. The two 27-billion-parameter models are substantial: the source estimates roughly 55 GB for each set of weights at 16-bit precision, meaning deployment requires data-center-class hardware rather than a typical gaming PC.

The benchmark results point to task-specific strengths rather than a universal lead. Nvidia reports Flora scoring 50.3 on the public WGO-Bench segmentation test, compared with Astra’s 46.0, and 37.4 versus 29.7 on the human-video subset of DenseCaptionBench. On the public RoboRewardBench, the paper reports Flora’s error was 27.8% lower than Astra’s. Those findings come from different tests and should not be treated as proof that one model is better for every robot-video workflow.

“research or evaluation purposes only”

— NVIDIA OneWay Noncommercial License

Production Costs and Commercial Use

The available figures do not establish Flora’s actual cost in a production deployment. The GPU rate is an assumption, throughput is modeled, and storage and overhead are omitted; Nvidia says the production GPU-hour denominator has not been verified. The benchmark also covers about 75 minutes of video, and the reported Astra total excludes attempts without usage data. It remains unclear how the costs would change across different video types, workloads, hardware arrangements or service requirements.

Quality beyond the reported evaluations is also uncertain. Nvidia conducted most of the tests described in the source material, and the results vary by dataset and task. No independent replication is provided here. The license wording restricts use to research or evaluation, while the source says Nvidia itself may use the models commercially; the scope of any such use is not detailed. Flora’s rewarder is a separate model and is not included in the 125-times captioning comparison.

Testing Beyond Nvidia’s Benchmarks

The next useful evidence would be independent evaluations on a broader mix of robot videos, with the same tasks and quality measures applied across Flora and competing systems. For the cost claim, readers would need measured throughput and full operating costs—including hardware utilization, storage and overhead—rather than a modeled GPU budget alone. Until those details are available, the estimate is best read as a benchmark-based indication of potential savings, not a quoted commercial service price.

Researchers can evaluate the released models under the stated license. Companies considering deployment will need to review the noncommercial terms and seek an appropriate commercial arrangement or another licensed option. Nvidia’s release makes the models and paper available now; no additional deployment milestone or independent cost study is specified in the source material.

Key Questions

What is Nvidia Flora?

Flora is a suite of three models for robot-video labeling: one finds action intervals, one captions them, and one scores steps or complete attempts.

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

Nvidia’s paper reports a ratio of about 125 to 1, based on $88.49 per video hour for Astra API usage and a modeled Flora cost of $0.72. The two figures are not like-for-like measured production costs, and the test covered about 75 minutes of video.

Can a company use Flora in a commercial product?

The stated NVIDIA OneWay Noncommercial License allows use for research or evaluation purposes. It does not grant ordinary commercial deployment rights; companies should review the license terms before use.

Does Flora outperform GPT-6 Astra on every test?

No. Nvidia reports Flora ahead on several measures, including human-video labeling and action segmentation, but says Astra did better on robot videos, caption accuracy and rubric scoring. Results vary by task and benchmark.

Source: ThorstenMeyerAI.com