Hierarchical Multiscale Diffuser for Extendable Long-Horizon Planning

Chang Chen, Hany Hamed, Doojin Baek, Yoshua Bengio, Sungjin Ahn

Keywords: Long-Horizon Planning, Diffusion, Hierarchical, Multiscale

TL;DR: We propose hierarchical multiscale planning method based on progressive trajectory extension and diffusion models.

Abstract: This paper introduces the Hierarchical Multiscale Diffuser (HM-Diffuser), a novel approach for efficient long-horizon planning. Building on recent advances in diffusion-based planning, our method addresses the challenge of planning over horizons significantly longer than those available in the training data. We decompose the problem into two key subproblems. The first phase, Progressive Trajectory Extension (PTE), involves stitching short trajectories together to create datasets with progressively longer trajectories. In the second phase, we train the HM-Diffuser on these extended datasets, preserving computational efficiency while enhancing long-horizon planning capabilities. The hierarchical structure of the HM-Diffuser allows for subgoal generation at multiple temporal resolutions, enabling a top-down planning approach that aligns high-level, long-term goals with low-level, short-term actions. Experimental results demonstrate that the combined PTE and HM-Diffuser approach effectively generates long-horizon plans, extending far beyond the originally provided trajectories.

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10313