Nvidia releases Flora open model suite for robot data annotation

Nvidia released Flora, an open model suite for robotics data annotation, on October 20, 2026, publishing the model weights on Hugging Face. According to aitimes.com, the suite is intended for research use and carries Nvidia's non-commercial licence.

Flora comprises three models: Flora-TempSeg 9B for segmenting actions in video, Flora-Captioner 27B for timestamped action captions, and Flora-Rewarder 27B for judging action quality and detecting failures. Nvidia says the suite can lower robotics data production costs by as much as 125 times.

The company reports that Flora can process up to 200,000 hours of video per day using 512 GB300 GPU servers. It estimates the cost at about 0.72 dollars per hour of source video, against about 88.49 dollars for GPT-6 Astra.

Benchmark results and GR00T training

In an internal evaluation of 50 episodes, Flora scored 57.1% versus 53.7% for GPT-6 Astra at maximum reasoning effort. On the RoboReward benchmark its mean absolute error was 40.0, compared with 55.4 for Astra, while failure detection recall on DenseRewardBench reached 53.3%.

Nvidia uses Flora to train GR00T, its humanoid robot foundation model. The company says Flora's outputs are data annotations and evaluations rather than direct robot control commands.

Until now, robotics data annotation has often been handled manually or through costly large multimodal model APIs, which has limited large-scale data collection.