Continuous 3D Perception Model with Persistent State

This seems like it could be useful for game engines—continuous 3D scene updates from RGB input sound a lot like dynamic world streaming. Could this be practical for real-time asset generation or scene reconstruction in games?