TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Nvidia has released Flora, an open set of models for labeling robot-training videos with action descriptions and performance ratings. Nvidia estimates Flora costs $0.72 per video hour, versus $88.49 for GPT-6 Astra, but those figures use an assumed GPU rental rate and the models perform differently across evaluation tests.
Nvidia has open-sourced Flora, a collection of models that labels robot-training videos with timestamped actions and ratings of task progress and success. Nvidia says the system can process video at an estimated $0.72 per hour, compared with $88.49 for GPT-6 Astra in the company’s test, a cost claim based on projected GPU rental prices rather than measured production costs.
Flora is designed to produce three kinds of annotation used in robotics training and monitoring: descriptions of actions and when they happen, step-by-step ratings of how well a robot performs, and judgments about whether an episode’s task succeeded. Nvidia’s GEAR research lab says many existing tools cover only one of those needs, while using general-purpose models to label large video collections can be expensive.
The system consists of three fine-tuned Qwen3.5 models. A 9-billion-parameter model identifies action segments, and a 27-billion-parameter model describes them using surrounding video context. A second 27-billion-parameter model, Flora-Rewarder, can rate overall task success from 1 to 5, classify half-second steps as normal, suboptimal or failed, and score episodes against a rubric supplied in the prompt. Nvidia says the models are available on Hugging Face.
Nvidia reports that Flora is already being used during training of its GR00T robotics models and can annotate up to 200,000 hours of video per day across 512 GB300 GPUs. That is a company-reported operating scale, not an independently verified performance measure. The company’s reported quality comparison used 50 videos: Flora scored 57.1% and GPT-6 Astra 53.7%, according to Nvidia.
Lowering the Cost of Robot Data
Robotics developers need large collections of labeled examples to train systems to recognize actions, learn from mistakes and judge whether a task has been completed. If Flora’s cost estimate holds in real use, automated labeling could make it less expensive to process more video and experiment with different training approaches, including transferring human demonstrations to robots.
The headline price comparison does not mean Flora is better at every task or that its actual operating cost is settled. Nvidia’s $0.72 figure assumes a rental price of $10 per GPU per hour; the report describes it as a projection, not a production measurement. Benchmark results also vary by task. Flora’s contribution may be most useful where action timing and detection of errors during an episode matter, rather than as a universal replacement for commercial models or human review.
The release also fits Nvidia’s wider robotics business: the company supplies hardware, simulation tools, training systems and models. An open labeling tool can support developers building on that stack while addressing a data-processing bottleneck. It gives researchers a model option they can inspect and test, but the available evidence does not establish how it performs across different robots, datasets or real-world deployments.
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 Fits Nvidia’s Robotics Stack
Robot-learning systems are trained on video as well as other data, but footage alone may not specify which action occurred at a given moment, how well it went, or whether the goal was achieved. Those labels can help select training examples, provide learning signals and monitor robot behavior. Flora combines video segmentation, captioning and reward scoring in one model suite.
Nvidia has previously presented automated data preparation and scoring as part of its robotics pipeline, and Flora adds detailed video labels and reward signals to that effort. The company has also positioned Flora as an open alternative to commercial services such as Scale AI’s Physical AI Data Engine. Google’s announced Gemini 4 Argon targets a related training-video task by identifying steps and flagging failed actions, according to the source report.
The benchmarks cited in Nvidia’s paper do not give Flora a clean sweep. On the public WGO-Bench, Flora was strongest at locating action boundaries but scored below Astra overall and below Gemini 3.8 Flash on description accuracy. On Nvidia’s DenseCaptionBench, Flora led for human-recorded videos but ranked behind Astra on robot videos. The researchers also acknowledged a methodological limitation: GPT-6 Astra evaluated descriptions while also being one of the models being tested.
“The company estimates that labeling one hour of video costs about $0.72 with Flora, compared with about $88.49 with GPT-6 Astra.”
— Nvidia
NVIDIA RTX GPU server for machine learning
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Costs and Performance Still Vary
The $0.72-per-hour estimate depends on an assumed GPU rental price, and the source report does not provide measured production costs or a breakdown of expenses beyond that assumption. Nvidia’s 50-video comparison is also a limited test; it does not establish results across the full range of robotics footage or deployment conditions.
Flora’s strengths and weaknesses depend on the evaluation. The researchers report better detection of action boundaries and mid-episode failures in some tests, but Astra leads on rubric-based scoring and trajectory ranking, and on robot-video captioning in one benchmark. Flora can also falsely flag normal steps as errors: the paper reports that it did so for about 11% of normal steps, while sometimes splitting longer actions too finely or misdescribing them.
In one reported test, Flora caught 53.3% of failed steps within an episode, while no competitor exceeded 11.7% and Astra caught none. The same research reports that Flora falsely flagged some normal steps, underscoring that reward labels may need review before they are used to train or assess a robot. How well these results transfer to other datasets and practical workflows remains unclear.
robotics development kit with camera and sensors
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Testing Flora in Robotics Workflows
Researchers and robotics developers can access the released models on Hugging Face and test them against their own video datasets. Further independent evaluations could clarify whether the projected cost advantage holds under actual hardware and staffing costs, and how performance changes across robot types, recording conditions and annotation tasks.
Nvidia’s stated use of Flora in GR00T training provides one indication of how the company plans to apply the system, but the report does not give a timetable for additional releases or independent trials. The next useful evidence will include real-world cost accounting, comparisons on more diverse datasets and results showing whether automated labels improve robot training and monitoring without adding unacceptable errors.
high speed NVMe SSD for AI video datasets
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Nvidia Flora?
Flora is an open set of fine-tuned Qwen3.5 models that segments robot videos, describes actions and scores task progress or outcomes. Nvidia says the models are available on Hugging Face.
Is Flora proven to cost 125 times less than GPT-6 Astra?
No. Nvidia’s comparison estimates $0.72 per video hour for Flora and $88.49 for Astra, but the Flora figure assumes a $10 GPU rental cost per hour. The source describes the figures as projections, not measured production costs.
Does Flora outperform GPT-6 Astra on every benchmark?
No. Nvidia reports that Flora did better on some measures, including action-boundary detection and certain reward ratings, while Astra led on other measures, including rubric scoring and robot-video results on DenseCaptionBench.
What kinds of labels can Flora produce?
It can generate timestamped action descriptions, ratings of individual steps as normal, suboptimal or failed, and scores for overall task success or a user-defined rubric.
What should readers watch for next?
Independent tests and real-world cost reports can show whether Flora’s projected savings and benchmark results hold across varied robotics datasets and training workflows. The source does not provide a schedule for such evaluations.
Source: rss
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
