NVIDIA has introduced Flora, a suite of three artificial intelligence models designed to turn videos of robots and human gestures into training data. Already integrated into the GR00T pipeline, it automates the identification, description and evaluation of actions.
Presented on 20 October by NVIDIA’s GEAR research team, Flora addresses an important step in robot development: preparing the videos used to train their AI models. A filmed demonstration has to be split into actions, given a description and evaluated to determine whether the movement was performed correctly. Flora automates these tasks to make large video collections easier to use.
How Flora works: three models to segment, describe and score actions
Flora relies on three specialised models, each handling one stage of video processing:
- Flora-TempSeg splits recordings into sequences that correspond to individual actions and identifies their time intervals.
- Flora-Captioner produces a natural-language description of the intended action in each sequence, taking the surrounding frames into account.
- Flora-Rewarder evaluates the quality of the execution and assigns scores to a complete attempt or to some of its steps.
According to researcher Yunze Man, this processing chain works with videos showing human hands, robotic grippers and dexterous robotic hands. The timestamped descriptions can be used, among other things, to filter and balance training datasets.
How Flora separates an attempted action from its success
One of Flora’s key functions is to separate the intent of a movement from its actual result. When a robot tries to grasp an object and fails, Flora-Captioner can describe the intended action, while Flora-Rewarder evaluates its execution separately.
Generative Bionics and AMD plan AI data factory in Italy to train Gene.01 humanoid robots
This distinction prevents a failure from being treated as a successful demonstration when training data is prepared.
According to the Flora-Rewarder model card, overall evaluations use a 1 to 5 scale to measure the success of an attempt. The system can also distinguish between normal actions, suboptimal executions and failures. It additionally supports evaluation based on a sequence of task-specific criteria.
These annotations can be used to measure robot performance and to prepare additional training phases using reinforcement learning.
NVIDIA says Flora processes 200,000 hours of video per day
NVIDIA states that Flora has already processed 200,000 hours of video per day using 512 GB300 GPUs. This figure refers to the volume of video annotated each day on this hardware configuration, not to new hours of robot experience or training.
Enact Intelligence raises $30 million for robots that “feel” resistance during assembly
This stated capacity relies on substantial computing infrastructure. It reflects NVIDIA’s goal: automating the annotation of large amounts of video to feed robot learning systems.
Flora already supports NVIDIA’s GR00T training pipeline. The data it produces can help with selecting demonstrations and evaluating behaviours during the development of robot models.
Flora vs GPT-6 Astra: robot annotation performance and cost
NVIDIA claims that Flora delivers annotation quality competitive with GPT-6 Astra at roughly 125 times lower cost on the robot annotation tasks evaluated by its team.
The published results vary depending on the metric, however. Flora scores higher on the joint F1 score for describing human videos and on the macro F1 used for dense reward evaluation. GPT-6 Astra keeps the lead on the joint F1 score for describing robot videos.
These comparisons cover specialised video annotation and evaluation tasks, not the general capabilities of the AI models.
Where to get Flora: the models are available on Hugging Face
NVIDIA has published a Flora collection on Hugging Face containing the three components: TempSeg, Captioner and Rewarder. The model cards for Captioner and Rewarder describe them as models derived from Qwen3.5-27B.
Developers can therefore access Flora’s specialised components to work on gesture segmentation, description generation and the evaluation of robot executions. The project also comes with a research paper, benchmark results and usage examples.