This paper introduces ID-WM, a hierarchical world model that separates state-space reachability from local control, enabling better long-horizon planning.
The research question addresses the challenge of long-horizon task planning for embodied agents, where existing hierarchical methods struggle with cumulative errors and search complexity.
The proposed approach, ID-WM, uses a high-level model to sample feasible state transitions and a low-level model to convert these transitions into precise actions, allowing for more flexible and accurate planning.
The authors report effectiveness across three navigation and manipulation tasks, showing improvements over state-of-the-art planners and significant gains in performance for long-horizon tasks.
About this brief
This article summarizes a public preprint abstract. Its findings have not been independently reproduced for this publication. Consult the paper for methods, evaluation scope and limitations. Automated briefs may contain errors; the original source takes precedence.