Compositional Diffusion with Guided Search for Long-Horizon Planning

TL;DR

CDGS addresses the core challenge of mode-averaging in compositional diffusion for long-horizon planning by embedding a guided, population-based search inside the diffusion denoising process. By combining iterative resampling across overlapping segments with likelihood-based pruning derived from DDIM inversion, CDGS selectively explores coherent sequences of local modes, yielding globally feasible plans. The approach demonstrates strong performance on robotic task planning and extends to long-horizon content generation, including panoramas and videos, without requiring long-horizon training data. The work provides a general, plug-and-play inference-time framework that enhances the reliability and coherence of long-horizon generative planning across domains, with reproducible experiments and open-source resources.

Abstract

Generative models have emerged as powerful tools for planning, with compositional approaches offering particular promise for modeling long-horizon task distributions by composing together local, modular generative models. This compositional paradigm spans diverse domains, from multi-step manipulation planning to panoramic image synthesis to long video generation. However, compositional generative models face a critical challenge: when local distributions are multimodal, existing composition methods average incompatible modes, producing plans that are neither locally feasible nor globally coherent. We propose Compositional Diffusion with Guided Search (CDGS), which addresses this mode averaging problem by embedding search directly within the diffusion denoising process. Our method explores diverse combinations of local modes through population-based sampling, prunes infeasible candidates using likelihood-based filtering, and enforces global consistency through iterative resampling between overlapping segments. CDGS matches oracle performance on seven robot manipulation tasks, outperforming baselines that lack compositionality or require long-horizon training data. The approach generalizes across domains, enabling coherent text-guided panoramic images and long videos through effective local-to-global message passing. More details: https://cdgsearch.github.io/

Figures (18)

• Figure 1: Compositional Diffusion with Guided Search (CDGS) composes short-horizon plan distributions to sample long-horizon goal-directed plans directly at inference. Unlike naïve compositional sampling, it explores diverse plans and filters locally inconsistent paths to avoid "mode averaging", yielding globally coherent plans.
• Figure 2: Applications of CDGS. (Top) Long horizon motion planning: CDGS discovers a valid multi-step plan to move the blue cube to the green cube’s original position via : (1) using the hook to pull blue cube in workspace, (2) displace the green cube to make space and (3) moving the blue cube to the target position. (Mid) CDGS generates coherent panoramic images. (Bottom) CDGS can stitch short clips to generate consistent, longer videos.
• Figure 3: Running example. (a) Consider a 1D-domain of $\{x_{1:7}\}$ variable distributions and $\{y_{1:6}\}$ feasible directed transitions between the variables. There are two feasible long-horizon plans from start ($x_1$) to goal ($x_7$): one through the top and one through the bottom. (b) in naive-composition, sampled plans may choose to start in the top and end at the bottom, or vice versa. When this happens, the intermediate models $\{y_{2:5}\}$ will average the modes of intermediate variables $\{x_{2:6}\}$ to satisfy both constraints, manifesting in infeasible transitions (red) (c) adding iterative resampling reduces the frequency of mode-averaging (d) adding pruning eliminates plans with infeasible $y$
• Figure 4: Compositional diffusion with Guided Search. At each denoising timestep, CDGS iteratively denoises a batch of noisy candidate global plans by (i) iterative resampling to propagate information through averaged scores at overlaps (blue) and (ii) pruning candidates with local inconsistencies based on the predicted clean samples (yellow). This process ensures all local plans align and belong to high-likelihood regions of $p(y)$, producing globally coherent plans.
• Figure 5: Left: Visualizing plan pruning. When compositional sampling chooses an infeasible mode sequence, the resulting plan can hallucinate out-of-distribution transitions due to mode-averaging as explained in \ref{['sec:method']}. For instance, (a) Infeasible transitions:inhand(hook) precondition is never met for place(hook), and (b) State hallucination:cube moves under(rack) as a result of averaging toward the goal state, despite being geometrically infeasible for push(cube, hook). Our pruning objective (\ref{['eq:pruning_objective']}) ensures only feasible plans during denoising, where all transitions are in-distribution with our short-horizon transition diffusion model. Right: Scaling analysis. (H-7) denotes performance averaged over tasks of horizon 7. (c) Task planning success improves with batch size, with larger gains from more resampling steps. (d) Motion planning success improves with resampling steps, but only when batch size is large enough