Compositional Diffusion Planning with Chunk Selection and Reachability Measure

Chanyoung Park; Youngjoong Kim; Jaesik Park

cs.ROOct 28, 2026

Compositional Diffusion Planning with Chunk Selection and Reachability Measure

Authors: Chanyoung Park, Youngjoong Kim, Jaesik Park

Organizations: Seoul National University

Abstract

Compositional diffusion planners have demonstrated remarkable success on long-horizon diffusion planning by sampling short trajectories, namely chunks, and composing them into a long trajectory. However, averaging overlapping chunks can induce the mode-averaging problem, resulting in discontinuous trajectories. Several works aim to address this problem, whereas a non-negligible portion of plans remains infeasible. In this work, we propose Selective Compositional Diffuser (SCD), handling this problem by measuring reachability. To this end, we present a self-supervised learning method for the reachability classifier. This is trained on offline data and generated chunks jointly to address potential distributional mismatch, while reducing the bias inherited from the generated data. Based on this, we construct a graph whose edge costs represent whether consecutive chunks are reachable, and traverse the graph to find effective yet efficient routes. Lastly, we generate bridging trajectories and refine the composed trajectory for smooth transitions. We evaluate our method on OGBench stitching tasks. It practically reduces incompatible mode compositions and the mode-averaging problem compared to the studied baselines. Across tasks, SCD improves success rates, reaching state-of-the-art results on HumanoidMaze-Large and Giant tasks.

Figures & tables

Panel of Figure 1 — Figure 1: Overview of the SCD pipeline. Double-dotted lines in Fig. 1a represent the initially generated plans. To find the executable plans, (i) we generate plans as in Fig. 1a , (ii) evaluate each state connection using our reachability classifier as in Fig. 1b , (iii) select a minimum-cost state sequence as in Fig. 1c , and (iv) build the bridge and refine the generated trajectory as in Fig. 1d .

Panel of Figure 2 — Figure 2: Failure cases in HumanoidMaze-Giant. The red color denotes infeasible trajectories and the blue color denotes feasible ones. Valid refers to the percentage of feasible trajectories. More samples are provided in App. D , Fig. 12 .

Figure 3: An example of graph construction. This example follows the setting of Fig. 1b .

Family	Method	PointMaze			AntMaze			HumanoidMaze			Avg.
Family	Method	Medium	Large	Giant	Medium	Large	Giant	Medium	Large	Giant	Avg.
Offline GCRL	GCBC	$23{\pm}18$	$7{\pm}5$	$0{\pm}0$	$45{\pm}11$	$3{\pm}3$	$0{\pm}0$	$29{\pm}5$	$6{\pm}3$	$0{\pm}0$	$12.56$
	GCIVL	$70{\pm}14$	$12{\pm}6$	$0{\pm}0$	$44{\pm}6$	$18{\pm}2$	$0{\pm}0$	$12{\pm}2$	$1{\pm}1$	$0{\pm}0$	$17.44$
	GCIQL	$21{\pm}9$	$31{\pm}2$	$0{\pm}0$	$29{\pm}6$	$7{\pm}2$	$0{\pm}0$	$12{\pm}3$	$0{\pm}0$	$0{\pm}0$	$11.11$
	QRL	$80{\pm}12$	$84{\pm}15$	$50{\pm}8$	$59{\pm}7$	$18{\pm}2$	$0{\pm}0$	$18{\pm}2$	$3{\pm}1$	$0{\pm}0$	$34.67$
	CRL	$0{\pm}1$	$0{\pm}0$	$0{\pm}0$	$53{\pm}6$	$11{\pm}2$	$0{\pm}0$	$71{\pm}3$	$6{\pm}1$	$0{\pm}0$	$15.67$

Table 1: Success rates (%) on OGBench Stitch maze tasks. Each environment provides 5 goal-conditioned tasks and we evaluate 20 episodes per task for three random seeds.

Panel of Figure 4 — Figure 4: Qualitative results of long-horizon planning in HumanoidMaze-Giant. SCD successfully finds feasible connections among the chunk candidates. See Fig. 13 for more results.

Method	01	02	03	04	05	Avg.
CD	$30$	$5$	$15$	$10$	$40$	$20$
CDGS	$0$	$30$	$0$	$0$	$45$	$15$
ECD	$45$	$45$	$20$	$25$	$50$	$37$
RCD	$25$	$30$	$40$	$10$	$60$	$33$
SCD (Ours)	$\mathbf{95}$	$\mathbf{80}$	$\mathbf{55}$	$\mathbf{75}$	$\mathbf{75}$	$\mathbf{76}$

Table 2: The ratio of geometrically feasible plans. For each HumanoidMaze-Giant task, we sample 20 plans and report the fraction whose trajectories are geometrically feasible, not passing through walls. The numerical entries in the header row denote the task IDs.

Figure 5: Adaptive plan length of SCD. Each point represents one evaluation task, and five tasks are evaluated for each maze size. The x-axis represents a geometric distance of the shortest path and the y-axis represents the length of the generated plan divided by the planner horizon (

H

). Horizontal dashed lines indicate the fixed number of chunks used by compositional planners. Error bars show the standard deviation over 20 plans.

Figure 6: Reachability heatmap comparison. $D_{\phi}(x,y,h_{o})$ is evaluated for five anchor states (circles) in AntMaze-Giant, under four combinations of the ablation study in Sec. 4.4 . Including generated states reduces overconfident predictions in regions that are unseen in the offline dataset. Density reweighting removes the start-state bias induced by nonuniform sampling. Together, the two components produce reachability maps that better respect walls and traversable corridors.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Diffusion model
Component	PointMaze	AntMaze	HumanoidMaze
Planner dim.	2D	2D	2D
U-Net base dim.	128	128	128
U-Net dim. multipliers	$(1,2,4,8)$	$(1,2,4,8)$	$(1,2,4,8)$
Boundary encoder	CNN	CNN	CNN
Boundary encoding dim.	256	256	256

Appendix

Table 3: Hyperparameters of the released CompDiffuser.

Setting	Value
Query encoder $Q_{\psi}$	3-layer MLP, width 512, SiLU
Key encoder $K_{\theta}$	3-layer MLP, width 512, SiLU
Lag encoding	Sinusoidal positional encoding, 64-dim.
Maximum lag $L_{\max}$	56 by default / 128 for HumanoidMaze
Lag sampling	Uniform $\mathcal{U}[1,L_{\max})$ for 75%, $L_{\max}$ for 25%
$k$ -NN reweighting	$k=50$ , $w_{\mathrm{min}}=10^{-8}$ , $w_{\mathrm{max}}=20$

Appendix

Table 4: Training settings for the reachability classifier.

Setting	Value
Baseline	CompDiffuser, sample chunks in parallel (without interleaving)
Scheduler	50-step DDIM ( Song et al., 2021 )
Minimum segment length $M$	48 by default / 80 for HumanoidMaze (i.e., $M=H-2h_{o}$ )
Boundary stride $S$	24 by default / 40 for HumanoidMaze
Discount factor $\gamma$	0.999
Shortest-path solver	Dijkstra’s algorithm (SciPy, CPU)

Appendix

Table 5: Default settings for candidate generation, graph search, and trajectory refinement. We present environmental settings such as the chunk length

H

and the bridge budget

h_{o}

in Tab. 6 .

Figure 7: Feasible ratio over wall tolerance. Feasible ratio of the plans in Tab. 2 as a function of the tolerance

\delta

. The dashed line marks

\delta=0.3

used in Tab. 2 .

Environment	Size	Max steps	$\bm{H}$	$\bm{h_{o}}$	$\bm{K}$	$\bm{N}$	Replan thr.	Max replans
PointMaze	Medium	1000	160	56	3	40	1	0
	Large	1000	160	56	6	40	1	0
	Giant	1000	160	56	8	40	1	10
AntMaze	Medium	1000	160	56	3	40	4	10
	Large	2000	160	56	6	40	4	10
	Giant	2000	160	56	9	40	4	10

Appendix

Table 6: Planning and evaluation settings per environment. Max steps denotes the episode limit.

H

denotes the chunk horizon,

h_{o}

the overlap length of the pretrained planner,

K

the number of chunks per candidate plan, and

N

the number of candidate plans. Replan thr. and Max replans denote the tracking threshold and the maximum number of replans. All values follow the released configuration of CompDiffuser ( Luo et al., 2025 ) .

Figure 8: Examples of the four infeasible pair types in the reachability test set (AntMaze-Giant). Each arrow points from a source (circle) to a target (star), and gray cells denote walls.

Table 7: Ablation of hyperparameters on AntMaze-Giant. Each subtable varies one hyperparameter. We report success rates (%) as mean

\pm

std across three seeds.

Figure 9: Top-down plan overview. SCD rollout in HumanoidMaze-Giant, where green circles mark the humanoid at intervals along its path from the start state (cyan) to the goal state (pink). The insets magnify the boxed region, showing the humanoid following through the plan.

Panel of Figure 10 — Figure 10: Samples of offline trajectory data , with each source (circle) connected to a target (star).

Panel of Figure 11 — Figure 11: Samples of generated trajectory data.

Panel of Figure 12 — Figure 12: Failure cases in HumanoidMaze-Giant. The red color denotes infeasible trajectories and the blue color denotes feasible ones. Valid refers to the percentage of feasible trajectories.

Panel of Figure 13 — Figure 13: Qualitative results of long-horizon planning in HumanoidMaze-Giant. The red color denotes infeasible parts and the blue color denotes feasible parts.

Diffusion-Based Planning

Diffusion-based planning generates diverse action sequences or trajectories by progressively denoising candidates conditioned on goals, scenes, and constraints.

See all 141 papers

Refining Compositional Diffusion for Reliable Long-Horizon Planning
May 4, 2026Kyowoon Lee, Yunhao Luo, Anh Tong +1Trajectory Generation Diffusion-Based Planning
KAIST · University of Michigan · Korea University +1
Energy-based Compositional Diffusion Planning
Jun 19, 2026Tao Sun, Utkarsh Aashu Mishra, Jiaxin Lu +2Trajectory Generation Diffusion-Based Planning
Stanford University, California, USA · Georgia Institute of Technology, Georgia, USA · University of Texas at Austin, Texas, USA
Compositional Visual Planning via Inference-Time Diffusion Scaling
Mar 3, 2026Yixin Zhang, Yunhao Luo, Utkarsh Aashu Mishra +3Diffusion-Based Planning Long-Horizon Planning
Georgia Institute of Technology · University of Michigan

May 4, 2026J/K move · Enter open · S save

Refining Compositional Diffusion for Reliable Long-Horizon Planning

Kyowoon Lee, Yunhao Luo, Anh Tong, Jaesik Choi

KAIST · University of Michigan · Korea University+1

Trajectory Generation Diffusion-Based Planning Diffusion Guidance Long-Horizon Planning

Open paper AskPDF arXiv

Compositional diffusion planning generates long-horizon trajectories by stitching together overlapping short-horizon segments through score composition. However, when local plan distributions are multimodal, existing compositional methods suffer from mode-averaging, where averaging incompatible local modes leads to plans that are neither locally feasible nor globally coherent. We propose Refining Compositional Diffusion (RCD), a training-free guidance method that steers compositional sampling toward high-density, globally coherent plans. RCD leverages the self-reconstruction error of a pretrained diffusion model as a proxy for the log-density of composed plans, combined with an overlap consistency term that enforces consistency at segment boundaries. We show that the combined guidance concentrates sampling on high-density plans that mitigate mode-averaging. Experiments on challenging long-horizon tasks from OGBench, including locomotion, object manipulation, and pixel-based observations, demonstrate that RCD consistently outperforms existing methods.

Abstract

Diffusion-Based Planning

Explore similar work

Refining Compositional Diffusion for Reliable Long-Horizon Planning

Energy-based Compositional Diffusion Planning

Compositional Visual Planning via Inference-Time Diffusion Scaling