Key Info

Jim Fan announced the open-sourcing of Flora, a suite of open models designed to parse the physical world and convert raw footage into rich training signals for robot learning, claiming Astra-level performance at 125x lower cost.

Highlights

  • Flora already powers an in-house data engine labeling 200K hours of videos per day on 500 GPUs.
  • Provides high-quality, timestamped captions to curate and filter human egocentric and robot data.
  • Offers dense progress rewards for quality control and RL post-training.
  • Includes rubrics scoring for real-world benchmarking.
  • Model weights are released openly for public use.