World-Ego Modeling for Embodied Video Generation in Long-Horizon Navigation-Manipulation Tasks
The World-Ego Modeling paradigm decomposes video prediction into persistent world regularities and robot-centric ego dynamics to solve long-horizon degradation in hybrid navigation and manipulation tasks. The World-Ego Model (WEM) uses a vision-language state predictor to separate these components, employing a cascade-parallel mixture-of-experts generator with full disentanglement for superior performance on complex tasks.