Nvidia has released Flora, an AI model suite designed to help researchers label robot and human-action videos at a fraction of the estimated cost of using large proprietary AI models. Developed by Nvidia’s GEAR robotics research team, Flora can divide videos into action segments, generate timestamped descriptions and score how successfully a robot completes a task. Nvidia says its system delivers performance competitive with GPT-6 Astra at roughly 125 times lower cost.
The release addresses a major challenge in robotics: collecting video is only the beginning of training a capable robot. Researchers must also identify the actions taking place, determine whether those actions were effective and generate signals that help robots learn from success and failure. Nvidia says Flora already supports its internal GR00T robotics training pipeline, processing up to 200,000 hours of data per day across 512 GB300 GPUs. However, the widely reported cost advantage is an estimate based on specific hardware and pricing assumptions, and the released models carry a noncommercial license.
Key Takeaways
- New AI suite: Nvidia released Flora, a collection of models for annotating robot and human-action videos.
- 125x cost claim: Nvidia estimates that Flora can perform video annotation at approximately 1/125 the cost of GPT-6 Astra for a tested captioning workload.
- Three main functions: Flora segments actions, generates timestamped captions and evaluates task performance.
- Robotics training: The system is used in Nvidia’s internal GR00T robotics data pipeline.
- Large-scale processing: Nvidia reports labeling up to 200,000 hours of video daily across 512 GB300 GPUs.
- License limitation: The model cards specify noncommercial research and development use, meaning the release should not be interpreted as unrestricted commercial licensing.
What Is Nvidia Flora?
Flora is an AI data engine designed to turn raw video into structured information that can help train and evaluate robotics systems.
Robots learn from demonstrations, recorded attempts and repeated interactions with their environments. But a video showing a robot picking up an object does not automatically tell a training system when the grasp began, whether the movement was efficient or why the attempt failed.
Human reviewers can supply these labels, but manually annotating enormous video collections is time-consuming and expensive. General-purpose AI models can automate some of the work, although repeatedly sending video to proprietary model APIs can create substantial inference costs.
Flora is designed specifically for these robotics data tasks. Its models process footage and produce information that can be used to organize datasets, evaluate robot behavior and generate training signals.
Nvidia says Flora has already been used in its internal GR00T-2 training pipeline. By releasing the model weights and research materials, the company aims to give robotics researchers a way to build customized data-processing systems rather than relying entirely on proprietary APIs.
How Flora Works: Three Models for Robotics Data
Flora is not a single model that performs every task. It combines three specialized components, each addressing a different stage of video annotation.
| Model | Main function | Why it matters |
|---|---|---|
| Flora-TempSeg 9B | Identifies action boundaries and segments video into meaningful intervals | Helps organize long recordings into individual actions |
| Flora-Captioner 27B | Generates timestamped descriptions of actions | Converts visual demonstrations into language-based instructions |
| Flora-Rewarder 27B | Scores overall outcomes and individual actions | Helps assess success, identify failures and support robot learning |
The numbers in the model names indicate their approximate parameter counts in billions.
Flora-TempSeg: Identifying When Actions Happen
The first component identifies the boundaries between actions. A long recording might include a robot reaching toward an object, grasping it, lifting it and placing it somewhere else.
Instead of treating the entire recording as one event, the segmentation model identifies meaningful time intervals. These segments can help researchers organize datasets and align visual observations with individual actions.
Accurate segmentation is important because training systems need to associate the right instruction or reward with the correct portion of a demonstration.
Flora-Captioner: Turning Video Into Action Descriptions
The captioning model produces timestamped descriptions of the actions taking place in each segment.
For example, a video segment might be described as a robot moving its arm toward a cup or placing an object on a shelf. These descriptions can help researchers search, filter and balance large video datasets.
Nvidia’s evaluation reports strong results on some captioning benchmarks, particularly for human-recorded manipulation footage. However, the company’s published results also show that performance varies by benchmark and video type.
Flora-Rewarder: Evaluating Success and Failure
The reward model evaluates how well an action or an entire attempt went.
It can produce an overall outcome score, step-by-step action-quality scores and rubric-based evaluations. A robot training pipeline can use these signals to distinguish successful behavior from ineffective actions.
This matters for reinforcement learning, in which a system improves its behavior using feedback about the quality of its actions. Better reward annotations can help researchers evaluate policies and guide post-training.
Nvidia explicitly states that Flora’s outputs are annotations. The reward model is not intended to directly control robots or serve as the sole safety monitor for a physical system.
Why Nvidia Says Flora Is 125 Times Cheaper
The headline cost advantage is the most attention-grabbing part of the announcement.
Nvidia’s comparison puts Flora’s estimated annotation cost at approximately $0.72 per hour of source video, compared with approximately $88.49 per hour for GPT-6 Astra in the reported workload.
| Cost metric | Flora | GPT-6 Astra |
|---|---|---|
| Estimated cost per video hour | $0.72 | $88.49 |
| Relative cost | Baseline | Approximately 123 times higher |
| Cost basis | Estimated GPU compute | Reported model API usage |
| Main task in comparison | Video captioning | Video captioning |
<text color=”secondary” size=”sm”>Source: Nvidia’s Flora research and published comparison as reported by The Decoder. The figures are workload-specific estimates, not universal prices.</text>
The ratio of the displayed figures is about 123 times, consistent with Nvidia’s rounded claim of approximately 125 times lower cost.
There is an important caveat: Flora’s estimate assumes a rental price of $10 per GPU-hour and modeled throughput on eight GB300 GPUs. The figure is therefore a projection, not a measured production bill that every user can expect to reproduce. Actual costs will depend on hardware access, utilization, video characteristics, deployment overhead and engineering effort.
The comparison also focuses on a particular annotation workload. It does not establish that Flora is 125 times cheaper for every robotics task or that it outperforms GPT-6 Astra on every quality measure.
How Flora Compares With GPT-6 Astra
Nvidia reports that Flora is competitive with GPT-6 Astra on several of its internal evaluations and performs strongly on selected public benchmarks.
In a reported test involving 50 videos, Flora achieved an annotation-quality score of 57.1%, compared with 53.7% for GPT-6 Astra at full reasoning effort. That is a result in Nvidia’s evaluation, not an independent confirmation that Flora is generally superior.
The wider benchmark table is mixed. Nvidia’s results show Flora performing better on some human-video captioning and reward-model metrics, while GPT-6 Astra scores higher on certain robot-video captioning and rubric-evaluation measures.
This distinction matters because robotics data is not a single uniform problem. A model may identify human actions well but struggle with robot-specific movements, unusual camera views or subtle manipulation failures.
The practical takeaway is that Flora offers a specialized alternative worth evaluating. Teams should compare both annotation quality and total cost on their own data before choosing a system.
Why Video Labeling Matters for Humanoid Robots
Robotics developers increasingly use large datasets to train systems that can manipulate objects, navigate environments and perform physical tasks.
A foundation model for robotics needs examples of what successful behavior looks like, but it also needs ways to identify mistakes and measure progress. Video annotation helps convert raw demonstrations into structured training material.
Flora’s functions could support several workflows:
- Dataset curation: Identify useful action segments and filter recordings.
- Training-data descriptions: Generate searchable captions for demonstrations.
- Reinforcement learning: Provide action-quality and outcome signals for training.
- Robot evaluation: Compare performance across attempts and help identify failure patterns.
- Human-to-robot transfer: Organize human demonstrations that could help inform robot learning.
These capabilities may be useful to research groups working on humanoid robots, robotic manipulation and general-purpose physical AI. But the quality of the underlying video, the accuracy of the annotations and the way labels are incorporated into training all influence the final results.
Is Flora Truly Open Source?
Nvidia describes Flora as an open model suite and makes its model weights available through Hugging Face. That gives researchers the opportunity to inspect, download and run the models on compatible hardware rather than using only a hosted API.
However, open weights do not automatically mean unrestricted commercial use. The Flora-Rewarder model card specifies Nvidia’s OneWay Noncommercial License and says the model is intended for noncommercial robotics research and development, not commercial or production use. The relevant model cards should be checked individually before use.
That restriction matters for startups and businesses hoping to incorporate Flora into commercial robotics products or paid annotation services. They should review the applicable terms and obtain appropriate licensing if needed rather than assume that downloadable weights provide unrestricted commercial rights.
The distinction also affects how the cost comparison should be interpreted. Flora may be attractive for eligible research workloads, but the licensing terms are part of the real-world decision about whether an organization can deploy it.
The Bigger Picture
Flora highlights a growing trend in AI: specialized models can make particular parts of the development process cheaper without replacing general-purpose models across the board. In robotics, video annotation is a foundational task because it turns raw observations into structured information that can be used to train, evaluate and improve physical systems.
Nvidia’s release also strengthens its broader robotics strategy by making part of the data engine behind GR00T available to researchers. The combination of specialized models, large-scale GPU infrastructure and robotics software could help lower the cost of experimentation, particularly for groups that can run the models efficiently.
The next test is whether independent researchers can reproduce the reported cost and quality results on different datasets and hardware. Licensing is another key factor: broader adoption by commercial developers may depend on whether Nvidia offers terms suitable for production use.
FAQs
1. What is Nvidia Flora?
Flora is a suite of AI models designed to annotate robot and human-action videos. It segments actions, generates timestamped captions and scores the quality or outcome of tasks.
2. Is Flora really 125 times cheaper than GPT-6 Astra?
Nvidia estimates that Flora can cost roughly 125 times less for a specific video-captioning workload. The comparison uses a modeled GPU cost for Flora and reported API costs for GPT-6 Astra, so the result is not guaranteed for every task or deployment.
3. How does Flora help train robots?
It turns raw footage into action boundaries, descriptions and reward signals. Researchers can use these annotations to curate training data, evaluate robot behavior and support reinforcement-learning workflows.
4. Can businesses use Flora commercially?
The published Flora model cards specify noncommercial research and development use. Businesses should review the relevant license and obtain appropriate permissions before using the models in commercial or production systems.
5. Does Flora control robots directly?
No. Nvidia describes Flora as an annotation and evaluation tool. Its reward model is not intended to directly control robots or act as the sole safety monitor for a physical system.
Looking Ahead
Flora gives robotics researchers a specialized option for processing the large video datasets needed to train and evaluate increasingly capable robots. If independent evaluations confirm its cost and quality advantages, it could reduce the expense of building data pipelines and make experimentation more accessible to eligible research teams.
For now, the most important factors to watch are reproducibility, real-world compute costs, performance across different robot types and licensing terms. The release is a meaningful development in robotics infrastructure, but its long-term impact will depend on whether the reported advantages hold outside Nvidia’s own evaluation and internal production environment.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
