Preprint reports improved long-horizon planning with hierarchical world models

Researchers report that their ID-WM planning system improved performance by 38.5% over leading policy and diffusion-based planners across three navigation and manipulation suites. The preprint also reports a 135% performance gain over a single-level world model planner that uses model predictive control. The system separates choosing reachable intermediate goals from generating the precise actions needed to reach them. A high-level model proposes feasible transitions, while a low-level model translates those transitions into actions. The researchers demonstrate plans assembled from short training segments, including complex configurations unseen during training; the reported gains concern the evaluated task suites.