TL;DR of the paper, generated automatically.

Paper on world action models, which predict video and actions from language.

• Data: about 47,000 hours of general, egocentric, UMI and robot videos.
• Finding: mixing egocentric and UMI data improves control generalization.
• Results: 88.10% Easy, 61.56% Hard on RoboTwin; 40% real-world vs 23.33%.