Researchers Jim Fan and Yunze developed a data engine that assigns timestamped instructions and reward signals to raw video footage of humans and robots. This suite of models, called Flora, is now open-sourced by Nvidia and delivers performance comparable to Astra at 125x lower cost. The system currently processes 200,000 hours of video daily using 500 GPUs to curate training signals for machine learning.
Nvidia employs Flora to train its GR00T robotics project, using dense progress rewards for reinforcement learning post-training and rubric scoring for real-world benchmarks. By releasing the weights to the public, the company aims to help developers scale the frontier of robot learning through the automated filtering of egocentric and robot data.
