NVIDIA introduces Flora, an open model suite for labeling human and robot video

NVIDIA researchers introduced Flora, and Linxi “Jim” Fan announced on Oct. 20 that its model weights were being open-sourced. The suite is designed to produce captions and reward signals from video; NVIDIA says its data engine is used to train GR00T models. The company claims it labels 200,000 hours of video daily, while its cost and GPT-6 Astra performance comparisons come from NVIDIA's own evaluation.