acceptodds
Under review as a conference paper at ICLR 2027

Compositional Diffusion Planning with Chunk Selection and Reachability Measure

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.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Generative Model, Diffusion Planning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_vZ7oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work aims to improve CompDiffuser by addressing inconsistencies between generated trajectory chunks that can prevent them from being successfully merged in overlapping regions. To this end, the authors propose Selective Compositional Diffuser (SCD), which first uses CompDiffuser to generate multiple candidate plans. They then use a classifier to estimate reachability between candidate chunk boundaries and construct a graph based on these scores. Dijkstra’s algorithm selects a minimum-cost sequence of chunk segments, which are connected using diffusion-generated bridging trajectories. Finally, the assembled trajectory is refined. The experimental results demonstrate improvements over the original CompDiffuser.

Critical strengths
  • The overall idea of selecting and refining trajectories generated by CompDiffuser is interesting.

  • The experimental results demonstrate that the proposed method improves upon the original CompDiffuser.

Critical weaknesses
  • The overall novelty of the method appears limited, as its main contributions involve selecting and connecting trajectory segments generated by CompDiffuser. It remains unclear how much of the performance gain could be achieved by simply selecting the best complete trajectory from the same pool of CompDiffuser-generated candidates, without segment recombination or bridging.

  • Although the idea behind SCD is interesting, the motivation for its design requires further clarification. CompDiffuser has a straightforward rationale: decomposing long-horizon trajectory generation into overlapping chunks. For SCD, however, it is unclear why the newly generated bridges should be more reliable than the original overlapping regions. Could bridge generation introduce new infeasible transitions and compromise the connectivity of the selected segments? Additionally, could reducing CompDiffuser’s chunk horizon achieve similar improvements?

Questions for authors

Q1. Could the authors compare SCD with additional state-of-the-art diffusion-based planners, particularly hierarchical diffusion planners?

Q2. Could the authors also compare SCD with CompDiffuser using shorter chunk horizons to determine whether reducing the chunk horizon yields similar improvements?

2 comments

We thank the reviewer for their detailed and thoughtful feedback. We hope our responses below address the reviewer’s concerns.

W1. Gain over selecting the best complete trajectory

(W1) It remains unclear how much of the performance gain could be achieved by simply selecting the best complete trajectory from the same pool of CompDiffuser [1]-generated candidates, without segment recombination or bridging.

Following the reviewer's question, we compare SCD with two ways of selecting a single complete trajectory from the same pool of 40 CompDiffuser candidates on AntMaze-Giant: selection with the same classifier cost as SCD (the sum of −log⁡Dϕ-\log D_\phi along the trajectory), and an oracle upper bound, which executes all plans and selects the best performance.

Classifier-based selection improves over CompDiffuser’s own selection (65 ± 3 → 83 ± 2) but remains below SCD (85 ± 2). We also evaluate without replanning to observe the maximum score CompDiffuser could independently achieve with an optimal selector. There, even the oracle reaches only 31% success and yields 44% feasibility, whereas SCD reaches 39% and 88%. Selection is thus bounded by the best trajectory in the pool, whereas recombining segments across candidates can also produce feasible plans that no single candidate contains.

We also note that CompDiffuser reports 27 ± 4 without replanning. Since the evaluation details for this setting are not specified in the released configuration, we evaluate all methods in the same environment and with the same executor, so the relative comparison is unaffected.

Table. Success rates on PointMaze-Giant

MethodSuccess (w/ replan)Success (w/o replan)Feasibility
CompDiffuser68±350±147±1
Classifier-based selector70±262±153±2
Oracle selector–74±156±1
SCD91±178±353±4

Table. Success rates on AntMaze-Giant

MethodSuccess (w/ replan)Success (w/o replan)Feasibility
CompDiffuser65±318±228±2
Classifier-based selector83±226±144±3
Oracle selector–31±244±2
SCD85±239±388±4

*Oracle selector is not reported with replanning, since it is challenging to track all outcomes that depend on all subsequent replanned choices.

We reflected this comparison in the revised manuscript, Appendix C.3 (Table 14).

W2. Using bridges rather than overlap averaging

(W2) The motivation for its design requires further clarification. ... For SCD, however, it is unclear why the newly generated bridges should be more reliable than the original overlapping regions. Could bridge generation introduce new infeasible transitions and compromise the connectivity of the selected segments?

To compare the two, we kept the selected segments and trajectory refinement fixed and changed only the connection method. Bridges yield more feasible connections (98.0% vs. 87.8% before refinement on AntMaze-Giant), while this gap can be absorbed by the refinement (98.5% vs. 97.7% on AntMaze-Giant). Nonetheless, bridges still achieve a higher success rate (85% vs. 81% on AntMaze-Giant, 83% vs. 81% on HumanoidMaze-Giant). We reflected this comparison in the revised manuscript, Table 7 and Appendix C.2 (Table 13).

Q1. Comparison with hierarchical diffusion planners

(Q1) Could the authors compare SCD with additional state-of-the-art diffusion-based planners, particularly hierarchical diffusion planners?

In the official comment, O2, we compare SCD with graph-guided baselines (XDiffuser [2], C-MCTD [3], GAS [4]) and hierarchical diffusion planners (DSP[5] and HD[6]). SCD achieves the highest average success rate across the nine tasks (92.22), maintaining a success rate above 80 on all nine tasks. Please refer to this official comment, O2, for further details.

Q2. CompDiffuser with shorter chunk horizons

(Q2) Could the authors also compare SCD with CompDiffuser using shorter chunk horizons to determine whether reducing the chunk horizon yields similar improvements?

Following the reviewer’s suggestion, we retrained CompDiffuser on AntMaze-Giant with halved chunk length. The result supports that reducing chunk length does not improve performance on generating long-horizon trajectories.

Table. Success rates over chunk horizons

MethodChunk HorizonSuccess
CompDiffuser16065±3
CompDiffuser (retrained)8023±3
SCD16085±2

[1] CompDiffuser, NeurIPS 2025

[2] XDiffuser, Hassidof et al., 2026 (arXiv:2605.16863)

[3] C-MCTD, Yoon et al., NeurIPS 2025

[4] GAS, Baek et al., ICML 2025

[5] DSP, Zhant et al., 2026 (arXiv:2609.34575)

[6] HD, Chen et al., ICLR 2024

Dear Reviewer vZ7o,

We appreciate the time you spent on our paper. Below we list the parts of our response that relate to your comments:

  • W1: We compared SCD with selecting the best single trajectory from the same candidate pool. Oracle selector also stays below SCD.
  • W2: We compared bridges with averaging the overlaps, and bridges showed better results.
  • Q1 (O2): We added comparisons with graph-guided and recent diffusion planners, including a hierarchical one, and SCD achieves the highest average success rate.
  • Q2: We added CompDiffuser retrained with shorter chunks, which performs worse than the original.

Please let us know if there is any questions. We would be happy to discuss any remaining concerns before the discussion period ends.

Reviewer_cYGyOpenReview ↗
Rating
2(+1)
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes Selective Compositional Diffuser (SCD) to address mode averaging in compositional diffusion planning. SCD samples many candidate plans with a pretrained CompDiffuser model, scores connections between chunk boundary states with a learned reachability classifier, and runs Dijkstra over a chunk graph to select segments. It then generates bridges between the selected segments and refines the full trajectory. The classifier is trained contrastively, with negatives drawn from both offline and generated states and reweighted by a k-NN density estimate. On OGBench stitch tasks, SCD reaches an average success rate of 92.22%, compared with 90.11% for RCD, with the largest gain on HumanoidMaze-Giant (83% versus 67%).

Critical strengths
  • The paper is well written and easy to follow.
  • Mode averaging in compositional diffusion planning is an important problem.
  • The appendix provides enough implementation detail for reproduction.
Critical weaknesses
  1. The novelty is limited: SCD is an add-on to the released CompDiffuser, and its only new parts over existing reachability-graph planners are generated negatives and k-NN reweighting.
  2. Minimum-cost search over all cross-chunk bridges favors any false-positive shortcut and always returns a path, and 24% of SCD plans remain infeasible (Table 2).
  3. Reachability is estimated on 2D xy positions only (App. B.1), so it is unclear how the method extends beyond maze navigation.
  4. The closest graph-guided baselines, XDiffuser (arXiv:2605.16863), C-MCTD, or GAS, are not compared.
  5. No ablation measures task success without graph selection or without generated negatives and reweighting, and success depends strongly on γ (69.0% to 85.3%, Table 9).
Questions for authors

Please see the weaknesses above.

9 comments

We appreciate the reviewer's thoughtful and detailed feedback. We hope our responses below address the reviewer’s concerns.

W1. Novelty over existing reachability-graph planners

(W1) The novelty is limited: SCD is an add-on to the released CompDiffuser [1], and its only new parts over existing reachability-graph planners are generated negatives and k-NN reweighting.

Please note that existing graph-based planners such as GAS [2] and XDiffuser [3] connect subgoals with distance measures learned from offline datasets, which may become unreliable for states generated online, as our ablation study in Table 5 empirically shows. C-MCTD [4], on the other hand, expands search branches sequentially, generating and evaluating one branch at a time, which makes planning substantially slower.

We therefore construct graphs over the generated states and introduce generated negatives with k-NN reweighting. Since we use chunks generated in parallel directly for graph construction, we do not need additional guidance when generating candidate trajectories.

The comparison results in response to W4 support the effectiveness of our method in improving geometric feasibility with limited additional planning cost. Our method also achieves higher average success rates than other graph-based planning methods. We reflected these distinctions in Section 2 and in the summary of Section 3.4.

W2. False-positive shortcuts and infeasible plans

(W2) Minimum-cost search over all cross-chunk bridges favors any false-positive shortcut and always returns a path, and 24% of SCD plans remain infeasible (Table 2).

The non-negligible fraction of infeasible plans has two main causes, neither of which is a false-positive shortcut by the classifier. As inverse-dynamics failures dominate HumanoidMaze execution even for wall-free plans, we analyze AntMaze-Giant instead: 36 of 300 plans (12%) touch a wall, with 63 wall-touching segments in total.

In the first case, the two states are actually reachable, but the generated segment passes through a wall. This accounts for 41 of the 63 cases. When we regenerate paths between the same endpoints and control with the stronger expert policy (expert policy provided by OGBench [5]), we find an executable path for all of these pairs with enough samples, confirming that these pairs are reachable.

In the second case, the failure originates from a connection to which the classifier had already assigned a low score (D < 0.1). The minimum-cost search selected it because every start-to-goal path in the graph contained a segment with D < 0.1. This accounts for the remaining 22 cases. Across all 300 plans, no selected connection with D ≥ 0.1 was unreachable, so the classifier did not admit false-positive shortcuts.

The first issue can be addressed by checking the selected segments in short windows with the classifier and regenerating those that fail, and the second by filtering before the graph search and adding samples until such a path is found, which we leave for future work.

We reflected this analysis in the revised manuscript, Section 4.1 and Appendix C.4.

W3. Extension beyond 2D maze navigation

(W3) Reachability is estimated on 2D xy positions only (App. B.1), so it is unclear how the method extends beyond maze navigation.

In response to the reviewer's question, we conducted additional experiments in higher-dimensional or non-maze tasks, and the proposed method is not restricted to a particular state dimensionality or to maze tasks.

We evaluated our method on higher-dimensional AntMaze-L tasks, which additionally encode the agent's orientation, joint configuration, and velocity. For the non-maze tasks, we test AntSoccer (17D, comprising the agent and ball states), in which the agent must dribble a ball to a target location, and Scene (14D), a robotic manipulation task.

Table. Success rates on high-dimensional tasks and non-maze tasks.

MethodAntMaze-L 15DAntMaze-L 29DAntSoccer Arena 17DAntSoccer Medium 17DScene Play 14D
GCBC [6]3±33±334±42±15±1
GCIVL [7]18±218±221±31±042±4
GCIQL [7]7±27±25±20±051±4
QRL [8]18±218±22±10±05±1
CRL [9]11±211±22±10±019±2
HIQL [10]67±567±523±28±238±3
CD62±164±469±317±342±3
CDGS [11]16±132±369±118±251±2
ECD [12]75±561±466±220±246±3
RCD [13]59±664±372±320±244±4
SCD89±181±381±123±157±1

Across all five tasks, SCD consistently improves upon CD with the same backbone and attains the highest success rate among the compositional diffusion methods considered. These results indicate that the reachability classifier and the graph-based selection extend naturally to high-dimensional state spaces and to tasks beyond maze navigation. We reflected these results in the revised manuscript, Section 4.3 (Table 3), with the evaluation details in Appendices B.1 and B.2.

For evaluation protocols, we briefly note the following.

  • For AntMaze-Large, results of GCBC–HIQL are taken from OGBench [5] (Table 2). Since ECD and CompDiffuser report their 15D results with different step budgets (10000 and 2000 steps, respectively), we re-evaluate CD, ECD, and SCD on the 15D task under the same 1000-step budget as OGBench. We run CDGS and RCD on both tasks and ECD on 29D with their public code on the same planner as SCD.
  • For AntSoccer, baseline results are taken from CompDiffuser Table 3 (GCBC–HIQL and CD), and the others from their respective papers.
  • For Scene, we use the planner and inverse dynamics model released by CDGS as the backbone, and evaluate CD and SCD under the same protocol. Results for CDGS are taken from the respective paper. ECD and RCD are evaluated under the same protocol, and results for GCBC–HIQL are also taken from the CDGS paper.

W4. Comparison with graph-guided baselines

(W4) The closest graph-guided baselines, XDiffuser (arXiv:2605.16863), C-MCTD, or GAS, are not compared.

In the official comment, O2, we compare SCD with graph-guided baselines (XDiffuser, C-MCTD, GAS). SCD achieves the highest average success rate across the nine tasks (92.22), maintaining a success rate above 80 on all nine tasks. Please refer to this official comment, O2, for further details. We reflected this comparison in the revised manuscript, Table 2.

W5. Ablations and sensitivity to γ\gamma

(W5) No ablation measures task success without graph selection or without generated negatives and reweighting, and success depends strongly on γ\gamma (69.0% to 85.3%, Table 9).

We address the three points in turn. We have included the following experiments in the revised manuscript.

Effect of graph selection. To isolate the contribution of graph selection, we kept the same 40 candidate plans and chose the one with the best classifier score, without recombining segments across candidates. On AntMaze-Giant, this variant attains 77 ± 4, compared with 65 ± 3 for CompDiffuser and 85 ± 2 for SCD. The classifier alone thus improves over CompDiffuser by selecting better candidates.

Graph selection improves further by composing compatible segments from different candidates, which mitigates infeasible plans more effectively than selecting a single candidate. Both components therefore contribute to the final performance. We reflected these experiments in the revised manuscript, Tables 6 and 7 (Section 4.5).

Effect of generated negatives and reweighting. In the official comment, O1, we provide results of final success rates for different classifier variants. The results show the same trend that generated negatives and reweighting improve final success rates, not only the feasibility scores (AntMaze: 11% → 85%, HumanoidMaze: 2% → 83%). Please refer to this official comment, O1, for further details.

Sensitivity to γ\gamma. Performance is robust to the per-step penalty −log⁡γ−\log \gamma over an order of magnitude (0.999 and 0.9999), although it degrades when the penalty is zero or too large. In practice, we use the same γ=0.999\gamma = 0.999 across all nine benchmark tasks and additional evaluations.

Table 12 shows similar success rates at γ=0.999\gamma = 0.999 and 0.9999 (85.3% and 83.7%). Note that the per-step penalty −log⁡γ-\log \gamma differs by a factor of 10 between these two values. Performance starts to drop at larger or smaller γ\gamma. Without the length penalty (γ=1\gamma=1), the search ignores plan length and may choose long detours. With a strong penalty (small γ\gamma), the search accepts risky shortcuts to shorten the plan. For example, at γ=0.99\gamma = 0.99, a shortcut that saves about 230 steps is taken even if the classifier scores it at only D=0.1D=0.1, whereas at γ=0.999\gamma=0.999 it would need to save about 2,300 steps. We reflected this point in the revised manuscript, Appendix C.1.


[1] Generative Trajectory Stitching through Diffusion Composition, Luo et al., NeurIPS 2025

[2] Graph-assisted stitching for offline hierarchical reinforcement learning, Baek et al., ICML 2025

[3] Plan First, Diffuse Later: Extrinsic Graph Guidance for Long-Horizon Diffusion Planning, Hassidof et al., 2026 (arXiv:2605.16863)

[4] Compositional Monte Carlo Tree Diffusion for Extendable Planning, Yoon et al., NeurIPS 2025

[5] OGBench: Benchmarking Offline Goal-Conditioned RL, Park et al., ICLR 2025

[6] Learning Latent Plans from Play, Lynch et al., CoRL 2020

[7] Offline Reinforcement Learning with Implicit Q-Learning, Kostrikov et al., ICLR 2022

[8] Optimal Goal-Reaching Reinforcement Learning via Quasimetric Learning, Wang et al., ICML 2023

[9] Contrastive Learning as Goal-Conditioned Reinforcement Learning, Eysenbach et al., NeurIPS 2022

[10] Offline Goal-Conditioned RL with Latent States as Actions, Park et al., NeurIPS 2023

[11] Compositional Diffusion with Guided Search for Long-Horizon Planning, Mishra et al., ICLR 2026

[12] Energy-based Compositional Diffusion Planning, Sun et al., ICML 2026

[13] Refining Compositional Diffusion for Reliable Long-Horizon Planning, Lee et al., NeurIPS 2026

Reviewer_sKkfOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Selective Compositional Diffuser (SCD) addresses incompatible trajectory chunks in compositional diffusion planning. It trains a reachability classifier using offline transitions and generated states, with density-based negative-sample reweighting. A graph search selects subsegments from candidate chunks; bridges are generated between selected segments, followed by trajectory refinement. Evaluation on nine OGBench stitching environments includes success rates, geometric feasibility, timing, classifier ablations, and hyperparameter studies. The largest gains occur on the difficult HumanoidMaze tasks.

Critical strengths
  1. Selecting compatible segments is a well-motivated response to mode averaging. The empirical improvement on difficult environments is substantial: HumanoidMaze-Large and Giant reach 89±2% and 83±3%, compared with the best listed baseline results of 79±3% and 67±4% (Table 1, p. 7). These results support the practical value of the approach even though it is not best on every environment. The reported deviations are across planning-sampling seeds, not independent retraining of the complete system (Appendix B.3.2, p. 16).

  2. The evaluation provides evidence beyond task success. On HumanoidMaze-Giant, geometric feasibility is 76% on average (Table 2, p. 7), and the shared-model timing comparison reports 9.08 seconds for SCD versus 11.07 seconds for RCD (Table 3, p. 8). The wall-tolerance analysis and explicit execution protocol (p. 16), and parameter studies (Table 9, p. 18), further support a useful quality–cost trade-off.

Critical weaknesses
  1. The ideal executability argument needs to be distinguished from the behavior of the implemented planner. Equation (11) is written for the optimal classifier D*, and p. 6 explicitly assumes a planner that can generate feasible local plans between reachable pairs within the time budget before interpreting the chain product. The paper also states that selection does not guarantee stable motion and introduces refinement (Section 3.4, p. 6). I therefore do not claim that the paper gives no assumptions or promises unconditional physical execution. The narrower concern is whether the finite learned D_phi and the product of its pairwise scores can be interpreted as conservative success probabilities for the actual bridge generator and controller, rather than useful graph-search scores. All planning operates in xy, while Ant/Humanoid execution uses a full-observation controller (pp. 14–16). Pairwise xy scores do not by themselves establish independence, sufficient state, or probability calibration for actual executions. This is a scope/interpretation concern, not a demonstrated failure of Dijkstra or a refutation of the ideal support argument.

  2. The classifier ablation does not directly establish the contribution to final task success. Table 4 (p. 9; protocol on p. 17) already provides the generated-negative/reweighting combinations and a k-neighbor sweep, and Table 9 (p. 18) reports downstream hyperparameter ablations. The missing connection is final planning success with generated negatives and reweighting individually disabled while holding candidate chunks fixed. A controlled classifier-ranking test does not fully characterize errors on edges selected by search; the authors themselves acknowledge this risk in the limitations (p. 9). Separately, Table 1 uses published baseline scores, while feasibility and timing are rerun (Appendix B.2, p. 16). The manuscript does specify shared checkpoints, controllers, and CompDiffuser-based replanning settings (pp. 14, 16–17); I have not established an actual budget mismatch. An explicit confirmation that these settings match the strongest borrowed baseline would remove the remaining comparison uncertainty.

Questions for authors
  1. Given the ideal-local-planner assumption stated on p. 6, is the product used only as a graph-search proxy, or is it intended as a calibrated lower bound on success of the implemented generator/controller? If the latter, what supports transferring the D* argument to D_phi and composing execution probabilities from xy scores?

  2. With candidate chunks and the controller held fixed, how do the four generated-negative/reweighting variants change final success and selected-bridge failure rates?

  3. Can the authors explicitly confirm that the strongest borrowed Table 1 comparison uses the same candidate budget and replanning settings documented in Table 8? I am asking for a comparison cross-check, not claiming that these settings are absent or already known to differ.

2 comments

We thank the reviewer for a careful review. We have incorporated the feedback and clarifications into the revised manuscript. We hope our responses address the reviewer's concerns.

W1, Q1: Reachability score as a graph-search proxy.

(Q1) is the product (Eq. 11) used only as a graph-search proxy, or is it intended as a calibrated lower bound on success of the implemented generator/controller?

Please note that the product in Eq. 11 is intended as a graph-search proxy, and it can be intuitively interpreted as a lower bound on executability (but not in the sense of an exact lower bound) in an idealized setting where the local planner can realize reachable pairs within the given time budget. We tempered this wording in Section 3.2: “As an intuitive explanation, we can interpret this as a lower bound on local executability.”

(W1) Pairwise xy scores do not by themselves establish independence, sufficient state, or probability calibration for actual executions.

We agree that the product of pairwise scores itself is not a conservative success probability, as pairwise xy scores do not establish a sufficient state or probability calibration for actual executions in the reported experimental settings. However, given a sufficient state, which the Markov assumption requires, it remains a lower bound even without independence between connections:

p(E∣A) ≥ ∏j=0J−1p(E∣vj,vj+1,hj) ≥ ∏j=0J−1D∗(vj,vj+1,hj).p(E\mid A)\ \ge\ \prod_{j=0}^{J-1} p(E\mid v_j, v_{j+1}, h_j)\ \ge\ \prod_{j=0}^{J-1} D^*(v_j, v_{j+1}, h_j).

The first inequality holds with equality under independence. It still holds even without independence, since the outcome of each connection depends only on its own endpoints under the Markov assumption. The second inequality follows from Eq. 11. Thus, with an ideal local planner, a well-trained discriminator, and a sufficient state, our proxy score can be interpreted as a conservative success probability, and remains meaningful.

We reflected this distinction in the revised manuscript, Sections 3.2 and 3.3, where the score is used as a proxy for executability. Furthermore, in response to cYGy-W3, we apply our approach to the full 29D AntMaze-Large observation, which includes the joint configuration used for actual execution. SCD achieves 81 ± 3 compared with 64 ± 4 for CompDiffuser, demonstrating its applicability to sufficient state settings.

W2, Q2: Final success rates of four ablation settings.

(Q2) How do the four generated-negative/reweighting variants change final success and selected-bridge failure rates?

In the official comment, O1, we provide results of final success rates for different classifier variants. The results show the same trend that generated negatives and reweighting improve final success rates, not only the feasibility scores (AntMaze: 11% → 85%, HumanoidMaze: 2% → 83%). Please refer to this official comment, O1, for further details.

W2, Q3: Clarification of experimental settings of Table 8

Can the authors explicitly confirm that the strongest borrowed Table 1 comparison uses the same candidate budget and replanning settings documented in Table 8 (Table 11 in the revision)

The experimental settings in Table 8 (Table 11 in the revision) are the same as those used for the strongest baselines borrowed in Table 1 (Table 2 in the revision), including the number of candidates, the number of chunks per environment, the replanning threshold, and the maximum number of replans. We reflected these settings in Table 11 and Appendix B.3.2 of the revised manuscript.

Dear Reviewer sKkf,

Thank you again for your review. We would like to point you to the parts of our response that address your questions:

  • W1, Q1: We clarified the scope of Eq. 11 as a graph-search proxy and explained when it remains a lower bound. We also added results on the full-state AntMaze, where SCD still outperforms CompDiffuser.

  • Q2 (O1): We report final success rates of the four classifier variants, which show that both generated negatives and reweighting improve success.

  • Q3: We clarified that Table 8 uses the same settings as the borrowed baselines in Table 1.

We would be grateful if you could check whether these resolve your concerns, and please let us know if you have any other questions.

Comments on the paper

2 comments

We thank the reviewers for their careful reading and constructive suggestions. We clarify the scope of our lower-bound claim and experimental settings, and we distinguish contributions by providing additional results.

For the shared questions, we provide common responses below.

O1: Final success rates of four ablation settings.

(sKkf-Q2) How do the four generated-negative/reweighting variants change final success and selected-bridge failure rates?

(cYGy-W5) No ablation measures task success without graph selection or without generated negatives and reweighting

We evaluated final success rates on AntMaze-Giant and HumanoidMaze-Giant with four different classifier variants. We also checked feasibility using the same rule as in Table 1 and report the fraction of executable plans and the fraction of executable bridges among those selected.

Table. Final success rates and feasibility metrics of classifier variants.

EnvironmentGenerated neg.ReweightingSuccess (%)Feasible plansFeasible bridges
AntMaze✓✓85±291%99%
✓✗84±190%99%
✗✓55±257%95%
✗✗11±06%86%
HumanoidMaze✓✓83±282%98%
✓✗77±146%95%
✗✓6±20%67%
✗✗2±20%69%

Both generated negatives and reweighting affect the fraction of executable plans and the success rate. Without generated negatives, the classifier is not exposed to generated endpoints as negatives. We observe that it therefore judges infeasible shortcuts between them as reachable and selects bridges crossing walls. Without reweighting, the negatives are unevenly distributed across locations, which can lead the classifier to overestimate reachability in regions where negatives are scarce. The classifier using both components achieves the highest success rate in both environments. We included the success rates in Table 6 (Section 4.5) of the revised manuscript.

O2: Additional comparison: graph-based models and competitive planners

(cYGy-W4) The closest graph-guided baselines, XDiffuser (arXiv:2605.16863), C-MCTD, or GAS, are not compared.

(vZ7o-Q1) Could the authors compare SCD with additional state-of-the-art diffusion-based planners, particularly hierarchical diffusion planners?

We additionally compared SCD with graph-guided baselines: XDiffuser [1], GAS [2], and C-MCTD [4] on the nine OGBench [9] stitch tasks. We also include recent competitive diffusion-based planners that support stitching tasks (CoFi [5], DSP [6], and HD [7]), using the results reported in their papers.

Table. Success rates on OGBench stitch tasks (Italic: reproduced)

MethodPM-MPM-LPM-GAM-MAM-LAM-GHM-MHM-LHM-GAvg.
GAS91±991±487±598±196±188±496±281±282±390.00
C-MCTD100±0100±0100±094±994±975±1878±0 (94±3)88±2 (70±3)48±3 (67±4)90.22
XDiffuser100±0100±090±096±266±287±092±390±289±490.00
CoFi100±0100±096±297±288±285±3––––
DSP66±644±41±186±352±59±373±420±435±642.89
HD24±317±20±071±136±20±0––––
SCD100±0100±091±195±192±185±295±189±283±292.22

SCD achieves the highest average success rate across the nine tasks, maintaining a success rate above 80 on all nine tasks. Together with the structural differences discussed in cYGy-W1, these results indicate that SCD performs consistently well across environments. We added GAS, C-MCTD, XDiffuser, and DSP in Table 2 and their evaluation protocols in Appendix B.2 of the revised manuscript.

For evaluation protocols, we briefly note the following.

  • For GAS, AntMaze results are taken from the GAS paper (Table 1). HumanoidMaze results are taken from the GAS evaluation reported in TTGS [3] (Table 4). Since GAS provides no configuration for PointMaze, we trained it with the authors’ AntMaze settings and adjusted two evaluation settings for PointMaze.
  • For C-MCTD, PointMaze and AntMaze results are taken from its paper. We trained C-MCTD’s own planner with the official code and executed its plans with the same inverse dynamics model and replanning rule as SCD. Values in parentheses are obtained with hyperparameters tuned for HumanoidMaze. We choose the higher value for each task in the average.
  • For XDiffuser, we evaluated the public code under the same protocol as SCD: the step budgets and replanning settings, DDIM 50-step sampling, actions clipped to [-1, 1] as in OGBench PointMaze to prevent out-of-range actions that move the agent through walls. We also measure HumanoidMaze results, which are not reported in the paper.
  • For HD, we use the results reported from SCoTS [8].

[1] XDiffuser, Hassidof et al., 2026

[2] GAS, Baek et al., ICML 2025

[3] TTGS, Opryshko et al., ICML 2026

[4] C-MCTD, Yoon et al., NeurIPS 2025

[5] CoFi, Park et al., arXiv:2606.00837

[6] DSP, Zhang et al., arXiv:2609.34575

[7] HD, Chen et al., ICLR 2024

[8] SCoTS, Lee et al., NeurIPS 2025

[9] OGBench, Park et al., ICLR 2025

Dear Reviewers,

Less than two days remain before the discussion period ends on November 18 (AoE). If you have not yet replied to the author rebuttal, please do so as soon as possible, indicating which of your concerns are resolved and which remain.

Best regards, Area Chair

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.