Nvidia releases Flora models for annotating and evaluating robotics videos
The Flora models from Nvidia are available on Hugging Face. They segment robotics videos into actions, describe them and evaluate both task progress and success. The company estimates costs at 0.72 USD per hour of video; this is a projection, not a measurement of operating costs.
Nvidia has released the Flora models for creating temporal annotations of robotics videos and evaluating both task progress and outcomes. The models are available on Hugging Face. According to the company, they are already used in training GR00T models and annotate up to 200 000 hours of video per day on 512 GB300 GPUs.
The system consists of three modified Qwen3.5 models: a model with 9 billion parameters segments video into actions, while a model with 27 billion parameters describes them in the context of surrounding activity. Another model with 27 billion parameters, Flora-Rewarder, rates the success of the entire task on a scale of 1 to 5, classifies half-second steps and assigns scores based on specified evaluation criteria.
Nvidia estimates costs at 0.72 USD per hour of video, compared with 88.49 USD when using GPT-6 Astra. The calculation for the Flora models assumes GPU rental at 10 USD per hour; these are not measured operating costs. In a test involving 50 videos, the company says the Flora models achieved slightly better annotation quality than GPT-6 Astra.
However, results vary by task. According to the researchers, the Flora models excel particularly at identifying action boundaries, while GPT-6 Astra leads in the overall score on WGO-Bench and in evaluation against criteria. Flora-Rewarder detected 53.3 % of erroneous steps within tasks, but incorrectly flagged approximately 11 % of normal steps. The researchers also note that GPT-6 Astra evaluated descriptions in a test in which it was itself a participant.
Why it matters
Researchers can obtain descriptions from their own videos to help select training data, as well as evaluation signals for robot learning. However, planning large-scale processing requires verifying actual costs and error rates: neither the published cost estimate nor the results of individual tests guarantee the same benefit for every dataset.
Two audiences, two different impacts
What this means
For individuals
Developers of robotics models can use the released models to prepare their own data. When reviewing annotations, they should account for the fact that the system sometimes splits longer actions unnecessarily and makes errors in describing them.
For a business
For companies preparing training data for robots, this is an available option for automating annotation and task evaluation. However, budgets cannot be based solely on the estimate of 0.72 USD per hour of video, because it does not reflect measured operating costs, and annotation quality may require additional review.
ProcessesCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.