Skip to main content
Back to timeline
arXivSource publication:

MCDC changes only the initial noise: long-horizon stitching improves on OGBench and stacks with search-based planners

Synopsis

Compositional diffusion planners split a long-horizon task into short trajectory chunks that are generated separately and stitched, but independently initialized chunks can commit to incompatible local modes, yielding plans that are locally feasible yet globally incoherent. MCDC is a training-free inference method that replaces independent chunk noise with correlated Gaussian noise across neighboring chunks while keeping each chunk's marginal exactly standard Gaussian, biasing the sampler toward compatible overlap modes without touching the pretrained denoiser. On OGBench long-horizon Stitch tasks, MCDC improves consistently over independent sampling at nearly the same inference time and offers a favorable success-cost trade-off against search-based composition.

Source-provided article image: Marginal-Preserving Coupled Diffusion Composition for Long-Horizon Planning
arXiv

Interpretation

In long-horizon planning, incoherent stitched trajectories can originate in the sampling initialization rather than in a weak local denoiser. Compositional diffusion planners previously drew each trajectory chunk's initial noise independently, so when an overlap region admits several feasible passages, neighboring chunks could each pick a different mode. On AntMaze-Giant-Stitch, independently initialized chunks were each locally plausible yet produced wall-crossing connections or low-density transitions after stitching, which locates the problem in cross-chunk coordination.

MCDC replaces only the inference-time initial noise, letting neighboring trajectory chunks share part of their randomness. Each chunk's initial noise is written as a sum of a global shared component, an edge-shared component and a private component, normalized by the chunk's degree in the adjacency graph so that every chunk's marginal stays standard Gaussian. On OGBench PointMaze, AntMaze, HumanoidMaze and AntSoccer Stitch tasks, using the same pretrained local diffusion model as CompDiffuser, MCDC improves over independent sampling in all non-saturated settings with nearly unchanged inference time.

Perspective

The demonstrated scope is OGBench Stitch-style long-horizon tasks: PointMaze, AntMaze and HumanoidMaze at Medium, Large and Giant sizes, plus AntSoccer in Arena and Medium settings, all evaluated in simulation with state-trajectory planning.

Effectiveness is bounded by the quality of the pretrained local diffusion model: if the local denoiser cannot generate feasible chunks, or the required transition is not covered by the offline data, coupling the initial noises alone cannot recover the missing local capability. Coupling strengths are currently fixed inference-time hyperparameters, and different environments or task lengths may need different values; on PointMaze-Giant, raising the local strength from 0.70 to 0.85 lowered the mean success rate from 78.6% to 72.6%, and the source does not state the global weight at which that sweep was run.