Planning a long sequence of actions, like navigating a maze or assembling an object, is hard because small errors add up and the number of possible choices explodes. Existing methods often rely on pre-learned routines or have limited flexibility.
ID-WM solves this with two levels: a high-level model that only decides which nearby state to aim for next, and a low-level model that works out the exact actions to reach it. The high-level model uses a diffusion process, a technique that generates plausible states from random noise, to ensure the chosen next state is physically reachable.
In tests on navigation and manipulation tasks, ID-WM improved performance by 38.5% over leading planners. It also showed a 135% gain over single-level world models on long tasks, especially in new configurations, because errors don't compound as badly.