Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CHD: Coupled Hierarchical Diffusion for Long-Horizon Tasks

Ce Hao, Anxing Xiao, Zhiwei Xue, Harold Soh

TL;DR

CHD addresses long-horizon planning by tightly coupling high-level subgoal inference with low-level trajectory generation through a coupled diffusion framework. Building on a Joint Diffusion Model, CHD introduces a coupled hierarchical classifier, asynchronous parallel sampling, and segment-wise generation to satisfy bi-directional HL-LL coupling, parallelism, and horizon reduction. Empirical results across maze navigation, robot task planning, and real-robot demonstrations show CHD outperforms flat and existing hierarchical diffusion baselines while offering faster sampling than prior hierarchical methods. The approach enhances trajectory coherence, robustness, and scalability, enabling practical long-horizon planning in complex robotic tasks.

Abstract

Diffusion-based planners have shown strong performance in short-horizon tasks but often fail in complex, long-horizon settings. We trace the failure to loose coupling between high-level (HL) sub-goal selection and low-level (LL) trajectory generation, which leads to incoherent plans and degraded performance. We propose Coupled Hierarchical Diffusion (CHD), a framework that models HL sub-goals and LL trajectories jointly within a unified diffusion process. A shared classifier passes LL feedback upstream so that sub-goals self-correct while sampling proceeds. This tight HL-LL coupling improves trajectory coherence and enables scalable long-horizon diffusion planning. Experiments across maze navigation, tabletop manipulation, and household environments show that CHD consistently outperforms both flat and hierarchical diffusion baselines. Our website is: https://sites.google.com/view/chd2025/home

CHD: Coupled Hierarchical Diffusion for Long-Horizon Tasks

TL;DR

CHD addresses long-horizon planning by tightly coupling high-level subgoal inference with low-level trajectory generation through a coupled diffusion framework. Building on a Joint Diffusion Model, CHD introduces a coupled hierarchical classifier, asynchronous parallel sampling, and segment-wise generation to satisfy bi-directional HL-LL coupling, parallelism, and horizon reduction. Empirical results across maze navigation, robot task planning, and real-robot demonstrations show CHD outperforms flat and existing hierarchical diffusion baselines while offering faster sampling than prior hierarchical methods. The approach enhances trajectory coherence, robustness, and scalability, enabling practical long-horizon planning in complex robotic tasks.

Abstract

Diffusion-based planners have shown strong performance in short-horizon tasks but often fail in complex, long-horizon settings. We trace the failure to loose coupling between high-level (HL) sub-goal selection and low-level (LL) trajectory generation, which leads to incoherent plans and degraded performance. We propose Coupled Hierarchical Diffusion (CHD), a framework that models HL sub-goals and LL trajectories jointly within a unified diffusion process. A shared classifier passes LL feedback upstream so that sub-goals self-correct while sampling proceeds. This tight HL-LL coupling improves trajectory coherence and enables scalable long-horizon diffusion planning. Experiments across maze navigation, tabletop manipulation, and household environments show that CHD consistently outperforms both flat and hierarchical diffusion baselines. Our website is: https://sites.google.com/view/chd2025/home
Paper Structure (33 sections, 41 equations, 20 figures, 6 tables, 2 algorithms)

This paper contains 33 sections, 41 equations, 20 figures, 6 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Problem Formulation
  4. Diffusion Models as Planners
  5. Hierarchical Diffusion Planners
  6. Method: Coupled Hierarchical Diffusion
  7. Joint Diffusion Model (JDM)
  8. Coupled Hierarchical Diffusion Planner
  9. Experiments
  10. Maze Navigation
  11. Robot Task Planning
  12. Analysis and Discussion
  13. Real-Robot Demonstration
  14. Conclusion
  15. Related Works
  16. ...and 18 more sections

Figures (20)

  • Figure 1: Illustration of our Coupled Hierarchical Diffusion (CHD). Left: CHD generates the joint distribution of HL and LL through the denoising process. The HL subgoals may appear reasonable, but the resulting LL trajectories are sub-optimal. Right: With the coupled classifier, CHD enables LL feedback to refine sub-optimal HL subgoals, leading to improved coherence and performance.
  • Figure 1: Robot Task Planning Results
  • Figure 2: CHD overcomes key limitations in hierarchical diffusion planning. (a) BHD plans HL subgoals and LL trajectories separately, lacking feedback and parallelism. (b) JDM enables tight HL and LL coupling but requires full joint-space diffusion. (c) CHD introduces classifier-guided LL to HL feedback and supports asynchronous, parallel generation. (d) Segment-wise generation further reduces horizon and complexity via localized planning.
  • Figure 2: Comparison Against CHD Ablations
  • Figure 3: Long-horizon trajectory planning in maze navigation. Left: Comparision of planned trajectories, $\bigstar$ represents sub-goals. Right: Normalized rewards in Maze2D environments in D4RL. CHD results are calculated over 150 seeds.
  • ...and 15 more figures