Marginal-Preserving Coupled Diffusion Composition for Long-Horizon Planning
Organizations: School of Information and Control Engineering, China University of Mining and Technology · Center for Artificial Intelligence, University College London, London
Abstract
Compositional diffusion planners extend short-horizon trajectory models to long-horizon goal-conditioned planning by generating and stitching multiple local trajectory chunks. However, in multimodal environments, local plausibility is not sufficient for global consistency: independently initialized chunks may denoise toward incompatible overlap modes, producing plans that are individually feasible but incoherent after stitching. Existing remedies typically add search, guidance, pruning, or graph-based structure at inference time, increasing computational cost or relying on task-specific machinery. We instead investigate graph-structured dependence among chunk-wise initial diffusion noises as a lightweight coordination mechanism for compositional trajectory planning. We propose Marginal-Preserving Coupled Diffusion Composition (MCDC), a training-free inference method for pretrained compositional diffusion planners. Instead of sampling chunk noises independently, MCDC constructs coupled Gaussian initializations whose per-chunk marginals remain exactly standard Gaussian, as required by the pretrained denoiser, while neighboring chunks receive positively correlated noise components. This preserves the standard-Gaussian initialization marginal of each chunk while biasing adjacent chunks toward compatible overlap-mode choices. We prove exact marginal preservation and show, under a simple multimodal overlap model, that positive coupling increases the probability of compatible mode selection. On OGBench long-horizon stitch tasks, MCDC consistently improves over independent compositional diffusion sampling while retaining nearly the same inference time, and achieves a favorable success--cost trade-off compared with search-based composition methods.
Figures & tables
| Env | Type | Size | GCBC | GCIVL | GCIQL | HIQL | DD | CD | CDGS | MCDC |
| Pointmaze | Stitch | Medium | ||||||||
| Large | ||||||||||
| Giant | ||||||||||
| Antmaze | Stitch | Medium | ||||||||
| Large | ||||||||||
| Giant |
| Dataset | CD | Global | Local | Hybrid |
| Point-M | ||||
| Point-L | ||||
| Point-G | ||||
| Ant-M | ||||
| Ant-L | ||||
| Ant-G |
| Dataset | CD | MCDC | CDGS | CDGS w/ MCDC |
| Point-M | ||||
| Point-L | ||||
| Point-G | ||||
| Ant-M | ||||
| Ant-L | ||||
| Ant-G |
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
| Item | Setting |
| Number of chunks | |
| Number of sampled compositions | |
| Random seed | |
| Independent initialization | |
| Hybrid MCDC | |
| Local marginal metric | fraction of chunks assigned to the upper basin |
| Environment | Type | Size | Env Steps |
| Pointmaze | Stitch | Medium | 1000 |
| Large | 1000 | ||
| Giant | 1000 | ||
| Antmaze | Stitch | Medium | 1000 |
| Large | 1000 | ||
| Giant | 2000 |
| Parameter | PointMaze | AntMaze | HumanoidMaze | AntSoccer |
| Architecture | U-Net | U-Net | U-Net | DiT |
| Segment length | 160 | 160 | 336 | 160 |
| Overlap size | 56 | 56 | 128 | 56 |
| Diffusion steps | 512 | 512 | 512 | 512 |
| 0.8 | 0.8 | 0.8 | 0.7 | |
| 0.15 | 0.15 | 0.15 | 0.1 |
| Environment | Type | Size | # Segments |
| Pointmaze | Stitch | Medium | 3 |
| Large | 6 | ||
| Giant | 8 | ||
| Antmaze | Stitch | Medium | 3 |
| Large | 6 | ||
| Giant | 9 |
| Env | Size | CD | MCDC | ||
| ✗ | ✓ | ✗ | ✓ | ||
| pointmaze | Large | ||||
| Giant | |||||
| antmaze | Large | ||||
| Giant | |||||
| Average | |||||
| Method | Pointmaze-Giant | Antmaze-Giant | ||
| Success (%) | Time (s) | Success (%) | Time (s) | |
| CD | ||||
| CDGS | ||||
| MCDC | ||||
| Peak GPU Memory (GiB) | |||
| Method | Forward NFE | PointMaze-Giant | AntMaze-Giant |
| CD | 100 | 0.84 | 0.85 |
| MCDC | 100 | ||
| CDGS | 518 | 0.80 | 0.81 |
| CDGS + MCDC | 518 | ||
| Method | PointMaze-Giant | AntMaze-Giant |
| CD | 68.0 | 65.2 |
| CD + Filter | 72.4 | 73.6 |
| MCDC | 77.6 | 76.6 |
| MCDC + Filter | 81.0 | 84.8 |
| Dataset | ECD | RCD | MCDC | ECD + MCDC | RCD + MCDC |
| AntMaze-Large | |||||
| AntMaze-Giant | |||||
| HumanoidMaze-Large | |||||
| HumanoidMaze-Giant |
| Dataset | Method | All Chunks Valid (%) | Fusion-Induced Failures (%) | Fused Plan Valid (%) |
| PointMaze-Giant | CD | |||
| CDGS | ||||
| MCDC | ||||
| AntMaze-Giant | CD | |||
| CDGS | ||||
| MCDC |
Diffusion-Based Planning
Diffusion-based planning generates diverse action sequences or trajectories by progressively denoising candidates conditioned on goals, scenes, and constraints.
See all 149 papers