RealFeedApp 745.52M latent-action descriptors generated from 30.1K hours of video
The RynnVLA-Latent model extracts 745.52 million action descriptors from 30.1k hours of human, robot, and simulation videos using future feature prediction and camera motion decoupling.
- Field
- robotics
- What they did
- Developed a method to extract hidden 3D motions from videos of humans and robots to train AI controllers without needing to record every specific gesture.
- Why it matters
- This will allow scaling robot training to thousands of hours of video data from various sources, rather than collecting limited datasets for each robot type separately.
#rla#robotics#vision-language-action#latent-actions#flow-matching
