acceptodds
Under review as a conference paper at ICLR 2027

P3: Persistent Particle Planning for Constrained Diffusion Control

Abstract

Diffusion models provide expressive priors over trajectories, but adapting these priors to test-time constraints requires maintaining feasibility and consistency across successive control decisions. We introduce Persistent Particle Planning (P3), a sequential Monte Carlo framework for diffusion control that maintains a weighted population of candidate plans across replanning steps. At each control step, P3 shifts and partially re-noises the candidate trajectories, refines them under the latest observation, and uses constraint-aware weighting and resampling to select among alternative continuations without retraining the diffusion model. We consider denoising and replanning as one Feynman–Kac particle system and analyze it under an idealized repair. We prove that the re-noising depth controls how reliably a kept plan stays on its route, and that keeping a rare, well-separated route takes far fewer plans than rediscovering it by sampling from scratch. Experiments under multiple test-time constraint configurations show that population reuse reduces route switching and improves success without constraint violations. Because P3 refines earlier plans instead of redrawing them, it also needs fewer denoising iterations per replan. On maze-navigation tasks, it plans faster than both regenerated populations and methods that correct a single sampled plan by constrained optimization. Code and pretrained models are available at https://github.com/p3-username/p3-anon.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Diffusion models, constrained control, model predictive control, sequential Monte Carlo, particle methods, trajectory planning, warm-start planning, Feynman–Kac models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MNKCOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes Persistent Particle Planning (P3), a test-time constrained planning framework for pretrained diffusion models. Rather than regenerating candidate trajectories from Gaussian noise at every replanning step, P3 maintains a persistent population of weighted plans. At each step, the previous trajectories are shifted, partially re-noised, and denoised under the latest observation. Constraint-aware weighting and sequential Monte Carlo (SMC) resampling guide the selection of feasible trajectories, while optional particle rejuvenation introduces fresh candidates to maintain exploration.

The authors provide theoretical analyses of trajectory retention under partial re-noising, the advantages of preserving rare feasible routes, and the effects of particle reuse. Experiments on D3IL Avoiding, PointMaze, and AntMaze with test-time constraints demonstrate improvements in constraint satisfaction, trajectory consistency, and computational efficiency compared with several diffusion-based planning baselines.

Critical strengths
  • Theoretical analysis: The authors provide a theoretical framework for understanding how partial re-noising affects trajectory retention and how maintaining a population of candidate plans can improve the probability of preserving rare feasible solutions. In particular, the analysis characterizes the trade-off between preserving previously discovered trajectories and introducing sufficient noise for adaptation. Although the results rely on idealized assumptions, they offer useful insights into the behavior of persistent diffusion-based planners and help motivate the proposed algorithm.

  • Empirical improvements in constrained diffusion planning: The experiments demonstrate that retaining candidate trajectories across replanning steps can substantially improve performance, particularly when feasible routes are difficult to discover through independent sampling. For example, P3 achieves 99% clean success on the cluttered obstacle-avoidance configuration, compared with 48% for conventional SMC. The ablations further distinguish the contributions of particle persistence and constraint-aware selection. These results provide meaningful empirical evidence for the practical benefits of trajectory reuse in diffusion-based control.

Critical weaknesses
  • Limited algorithmic novelty relative to existing planning methods: The central idea of reusing previously generated trajectories across successive planning steps has been extensively investigated in model-predictive control and related planning frameworks. In particular, sampling-based MPC methods reuse and perturb previous control sequences, while Stein Variational MPC [1] maintains multiple candidate solutions to preserve multimodality during planning. More directly, Diffusion-ES [2] already employs partial re-noising and denoising to refine a population of diffusion-generated trajectories. Although P3 extends these ideas by maintaining constraint-weighted particles and their selection history across replanning steps, the individual components such as trajectory warm-starting, population-based optimization, SMC resampling, and partial diffusion refinement are largely established techniques.

  • Limited evaluation of diverse test-time planning constraints: The experimental evaluation primarily focuses on obstacle avoidance and navigation under spatial exclusion constraints. While the results are encouraging, these tasks do not fully establish the general applicability of P3 to more diverse and complex test-time planning problems. In particular, Tree-Guided Diffusion Planner (TDP) [3] studies a closely related setting of zero-shot test-time planning using pretrained diffusion models. Their tasks involve more diverse objectives and constraints. I strongly encourage the authors to evaluate P3 on these benchmarks and compare it against TDP. Such experiments would provide a more comprehensive assessment of P3's ability to handle different types of test-time constraints and establish its effectiveness beyond the relatively restricted obstacle-avoidance settings considered in the current paper.

[1]: Alexander Lambert et al. Stein Variational Model Predictive Control. 2020.

[2]: Brian Yang et al. Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving. 2024.

[3]: Hyeonseong Jeon et al. Tree-Guided Diffusion Planner. 2025.

Questions for authors
  • Q1. Novelty relative to population-based planning methods: Could you clarify the key algorithmic differences between P3 and existing approaches such as Diffusion-ES, which also employs population-based trajectory refinement through partial diffusion noising and denoising? In particular, have you considered comparing P3 against a warm-started Diffusion-ES baseline under the same computational budget? This would help isolate the benefits of maintaining SMC particle weights and selection histories across replanning steps.

  • Q2. Adaptability to changing constraints: While trajectory persistence improves consistency, could it also reduce adaptability when constraints change substantially between replanning steps? For example, previously feasible trajectories may become infeasible due to newly introduced obstacles or changing goal conditions. Could you clarify how P3 handles such scenarios and whether the particle rejuvenation mechanism is sufficient to prevent premature convergence to previously preferred trajectories?

4 comments

We thank the reviewer for the precise questions. Q1 asks for the one experiment that isolates our contribution, and we ran it. We have updated the manuscript, and the changes are marked in blue.


W1, Q1: novelty and warm-started Diffusion-ES

We respectfully disagree, in particular on Stein Variational MPC, and we built the requested baseline and added it to Table 1.

The individual operations exist in prior work, as the reviewer says. None of the cited methods selects during denoising on a weighted population that it keeps across control steps, which is what P3 does. We formalize this process and analyze, under idealized assumptions, when retaining a discovered rare route is easier than repeatedly rediscovering it. The new warm-start comparisons show that reuse accounts for substantial gains, and P3 adds to them where the feasible route is rare and long, 0.85 against 0.74 for a warm-started Diffusion-ES on Narrow and 0.62 against 0.38 on Turn.

The overlap with Stein Variational MPC is that both keep several candidates between control steps. It has no diffusion model, so it cannot sample a multimodal trajectory distribution learned from demonstrations. It moves its particles by Stein variational gradient steps, and its weights set only the starting distribution of the next step, while P3 carries each plan's weight into the next denoising pass, where reweighting and resampling continue to act on it (Table 21). Keeping several candidates is not what produces our result. Persistent best-of-NN in Table R1 keeps its population and selects only at the end of denoising, and it reaches 0.56 on Narrow and 0.37 on Turn against 0.85 and 0.62 for P3.

P3 differs from Diffusion-ES in three ways:

  • Selection: Diffusion-ES selects among finished plans between its mutation passes, while P3 reweights and resamples during denoising.
  • Population: the released Diffusion-ES planner redraws its population at every replanning call, while P3 keeps it.
  • Weights: Diffusion-ES carries no weights between calls, while P3 does.

A warm-started Diffusion-ES is the right control, and it was missing. Diffusion-ES suggests warm starting in one sentence of its Sec. 3.2, but we checked the released code and its planner restarts from noise at every replanning call, so no published warm variant exists. We therefore built it ourselves, label it as our construction, and match its compute to P3, so that it differs from P3 only in the first and third points. The warm variant starts from the previous population shifted, takes the same number of denoiser calls per particle as P3, and selects by exponential reweighting over the current step only, with no weights or ancestry carried forward.

Table R1: success without constraint violation on Table 1's eight configurations at matched budgets. Bold marks the best value in each column among the rows shown.

MethodBothClutt.LeftRightBroadNarrowBlockedTurn
SMC0.970.480.910.970.750.510.000.53
FK steering0.930.470.880.890.790.420.010.44
Diffusion-ES0.980.380.960.900.760.510.000.43
Diffusion-ES (warm)0.970.980.990.980.860.740.210.38
Persistent best-of-NN0.930.870.860.920.750.560.200.37
P3 (ours)0.990.990.940.980.890.850.210.62

Three readings follow. Warm starting accounts for most of the gain, as the warm row moves from 0.38 to 0.98 on Clutt. relative to cold Diffusion-ES. P3 keeps its further margin on Narrow and Turn. On the easy configurations P3 and the warm baseline are tied, and on Left the warm baseline is ahead. Per-step planning times on one GPU are 117, 93, 24 and 186 ms for P3 against 172, 86, 23 and 171 ms for the warm baseline on Clutt., Narrow, Blocked and Turn (Table 10).

To isolate the carried weights, we reset them to uniform at every control step at the same compute. On the large maze with a disc in a corridor, this lowers success from 0.94 to 0.78 when the disc slides and from 0.96 to 0.76 when it stays still (Table 17 in Appendix F). On Avoiding the same reset changes success by at most 0.03 (Table 16), so the carried weights matter on the large maze and not on Avoiding.

On the other prior work, Table 21 in Appendix L compares the operations of each method. MPPI keeps one control sequence and redraws its samples around it. FK steering discards its population after each sample. The one Feynman-Kac model spanning denoising and replanning, with weights and ancestry carried across control steps, is what we claim. Table R1 is the evidence for selection during denoising on a kept population, and the weight-reset test is the evidence for the carried weights on the large maze.

W2: diverse constraints and TDP

We ported TDP and added a TDP-style manipulation task. P3 is above TDP on all eight configurations, and on the new task it places as many blocks as TDP, with more stages without violation and about half the planning time.

We did both halves of this. We ported TDP from its official code and ran it on all eight configurations, where it reaches 0.94, 0.38, 0.90, 0.89, 0.76, 0.50, 0.01 and 0.42 success without constraint violation, below P3 on every column, at 149 ms per step against 117 for P3 on Clutt.

We also added KUKA sequential block placement, a task family from TDP, with a protection constraint that forbids disturbing the three inactive blocks. Against TDP itself, on ten paired scenes, P3 and TDP both place 1.10 of four blocks, with 1.10 stages without violation for P3 against 0.70 for TDP, and 17.94 s of planning per episode against 34.76 s. Table R2 compares the other methods on a separate set of 20 scenes, so its P3 row differs from the ten-scene numbers.

Table R2: KUKA, 20 paired scenes, one pretrained model. Any violation is the share of episodes with a protection violation. A stage without violation places its active block without a protection violation. NFE counts particle denoiser evaluations, matched within 3%. Bold marks the best value in each column.

MethodAny violationPlaced / 4Stages without violation / 4NFE (10310^3)
SafeDiffuser85%0.550.4572
Best-of-NN85%0.650.5072
FK steering75%0.750.7072
SMC70%1.101.0072
Persistent best-of-NN60%1.200.9570.4
P3 (ours)30%1.401.2070.4

P3 violates the protection constraint in 30% of episodes against 60% for the next best method, and places the most blocks. Partial placement is the normal level on this benchmark. TDP's paper reports 36.9 on its 100-point scale for this task, about 1.5 of four blocks.

Beyond obstacle avoidance, the revision adds a height ceiling on Walker2d locomotion, where P3 is above the roof on 3.1% of steps against 9.3% for SMC, at less than half the denoiser evaluations.


Q2: adaptability when constraints change

We tested this, and persistence does not prevent adaptation.

We added Appendix F. An obstacle appears on the current plan at a fixed step, or moves across the workspace from one of two onset times, and the change is never disclosed in advance.

Table R3: success without constraint violation under changing constraints on Avoiding. Bold marks the best value in each column.

MethodStillMoves from step 24Moves from step 40Appears on the plan
Projection0.310.330.360.33
SMC0.500.540.470.67
P3 (ours)0.990.960.740.78
P3, no fresh plans0.900.900.890.67

P3 stays above SMC in all four columns, by 0.11 to 0.49. P3 loses 0.03 when the obstacle starts moving at step 24, from 0.99 to 0.96, and 0.25 when the motion starts at step 40. Constraints are reevaluated at every call, so a carried plan that has become infeasible loses its weight and is resampled away.

On whether adding fresh plans is sufficient, they help in three of our four tests with a change and cost success in one. They raise success from 0.90 to 0.96 when the obstacle moves from step 24, from 0.67 to 0.75 on the trials both variants ran when it appears on the plan, and from 0.35 to 0.94 on a separate changing-obstacle study on PointMaze, at 122 against 8 denoiser evaluations per step (Table 17). When the motion begins late, at step 40, the variant without fresh plans reaches 0.89 against 0.74 for full P3. The revised Limitations state that with a fixed number of fresh plans per step P3 can lose success when the change comes late.

On the same PointMaze study, P3 reaches 0.94 against 0.69 for SMC and against 0.78 with its weights reset, a gap that varies across models. On Avoiding, resetting the weights changes success by at most 0.03, so the weights carried from before a change do not hold P3 on a blocked route.


Presentation

We clarified terminology and sentence structure.

We made targeted language edits in the revised manuscript. We replaced the non-standard terms, added short definitions, and split long sentences in Sections 3.3 and 3.4. The appendices, tables and figure labels use the same terminology. The revised wording is marked in blue.

Thank you for conducting experiments I requested. However, new results brought me additional concerns on this paper.

  • Algorithmic novelty: It seems a substantial portion of P3's improvements arises from trajectory warm-starting. While the additional improvements from persistent particle weighting and within-denoising resampling are meaningful, I regard the overall algorithmic contribution as relatively incremental.

  • Comparison with TDP: P3 and TDP achieve the same average task completion (1.10 out of four blocks) on only ten paired scenes. Although P3 improves constraint satisfaction and planning efficiency, I would appreciate clarification on the comparability of planning budgets and evaluation setup.

I acknowledge that the additional experiments are helpful, so I raised my rating. However, as long as my concern about limited algorithmic novelty remains, I am learning toward rejection.

We thank the reviewer for reading the new results and raising the rating. The PDF is updated again, with changes in blue.


Algorithmic novelty

No released population baseline keeps its plans across replanning calls. The paper shows how large the gain from keeping them is, when it appears and why, and P3 adds to it where feasible plans are rare.

We agree that warm starting is an established idea and that it accounts for a substantial part of the gain. Earlier diffusion planners warm-start one trajectory (Table 21). The population methods we compare with, FK steering, TDP and the released Diffusion-ES, all redraw their plans at every replanning call, and the warm-started control exists because we built it at the reviewer's request (Appendix B). The paper explains the gain from a kept population with one Feynman–Kac model over denoising and replanning. Theorem 1 quantifies route retention under exact conditional denoising at a fixed observation without shifting, and Proposition 1 compares retained acceptable plans with fresh draws.

Controlled comparisons show what P3's within-denoising selection and carried weights add:

  1. Warm starting alone is unreliable. Persistent best-of-N (shift, partial re-noise and denoise, selection only at the end) helps on Clutt. but is worse than cold Best-of-N on Walker2d (12.9% vs. 8.7% of steps above the roof) and on Turn (0.37 vs. 0.43).
  2. P3 selects during denoising and carries its weights. Diffusion-ES (warm) reweights and resamples once per pass from current-step scores. On Turn, both warm controls (0.38, 0.37) fall below cold SMC (0.53), while P3 reaches 0.62.

Table R4: persistent controls at matched denoiser budgets, on four of Table 1's eight configurations and the two new tasks. Bold marks the best value, and both values in the Walker2d tie.

MethodResamplingCarried weightsClutt.BroadNarrowTurnWalker2d above roofKUKA violationKUKA placed / 4
Persistent best-of-NNnoneno0.870.750.560.3712.9%60%1.20
Diffusion-ES (warm)after passno0.980.860.740.383.3%50%1.30
P3 (ours)in denoisingyes0.990.890.850.623.1%30%1.40

When an obstacle appears on the plan, P3 reaches 0.78 vs. 0.67 for Diffusion-ES (warm) (+0.11+0.11, 95% CI [0.04,0.17][0.04, 0.17]). On the other four Table 1 configurations the two are tied, except Left, where the warm control is ahead. On the other new tests, it ties P3 on Walker2d, is within 0.04 on Avoiding with a moving obstacle, and leads on the sliding-disc maze at 0.99 vs. 0.94 (Tables 2, 16, 17).


Comparison with TDP

Both methods ran on the same model, scenes, cost, grading and denoiser budget. P3 places twice as many blocks as TDP on the 20 scenes of Table 20, and 1.30 against 0.83 over all 30 scenes.

The setup is identical for both methods:

  • Model & scenes: the released TDP KUKA checkpoint, with the same initial block positions, placement order and targets.
  • Protocol & grading: four stages and eight planning calls per episode (replan after 64 commands), graded at every physics substep with TDP's placement test.
  • Cost: shared target error plus protection penalty; TDP's tree is scored by it, so TDP sees the constraint.
  • Budget: 70,400 particle denoiser evaluations per episode for both. P3 keeps 16 plans (four full calls of 1,000 steps, four warm calls of 100). TDP draws 8 plans (eight calls of 1,000 parent and 100 child steps) with its released settings.

Ten scenes were too few, as the reviewer notes. We ran TDP on the 20 scenes of Table 20, plus a second TDP row with the protection penalty gradient in its guidance.

Table R5: KUKA, 20 paired scenes, 70,400 denoiser evaluations per episode. Bold marks the best value.

MethodAny violationPlaced / 4Stages without violation / 4
TDP65%0.700.50
TDP, protection guidance60%0.900.55
P3 (ours)30%1.401.20

P3 places 1.40 blocks against 0.70 for TDP (paired difference +0.70+0.70, 95% CI [0.25,1.15][0.25, 1.15]) and lowers episode violations by 3535 percentage points (95% CI [−65,−5][-65, -5]). These 20 scenes are separate from the ten earlier ones. Over all 30, P3 places 1.30 blocks against 0.83, with a violation in 37% of episodes against 67%. Both TDP rows are in Table 20 (Appendix K). On an idle single GPU on the ten earlier scenes, planning took 17.94 s/episode for P3 vs. 34.76 s for TDP at equal denoiser calls, due to TDP's guidance and tree operations. On Walker2d, TDP with roof gradient guidance leaves 4.1% of steps above the roof vs. 3.1% for P3 at 2.7×2.7\times the denoiser calls; a 3×3\times stronger gradient lowers this to 0.1%, but the walker falls in 15 of 48 episodes (Table 19).

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

The paper introduce the persistent particle planning for the diffusion policy considering safety constraints. The framework includes a few components, including denoising a few trajectories from gaussian noise, using the constraints violations to compute the weights for reweighting and resampling, executing the first action, renoising the rest of the trajectories, denoising them again for shifting plan and executing the most safe one. Also, the theory is given for the effectiveness of the framework. Experiments on D3IL Avoiding, PointMaze and AntMaze show that the algorithm could generally get better clean success and less wall-clock time.

Critical strengths
  1. The idea is simple and clear, and each component is useful from the motivation and is easy to follow.
  2. Ablation experiments are conducted to make sure that every component is useful for the clean success of the tasks.
Critical weaknesses
  1. The contribution is not fully clear to the reviewer. To my understanding, the paper leverages some existing components in the diffusion model to solve the control problem considering safety. The author should either propose new ideas in diffusion policy and verify in the toy examples or study a few techniques on the difficult tasks. However, the experiments in this paper are not complicated tasks where the state is low-dimensional and the exact formulation of the constraints are known.

  2. It is not clear that how the method applies to real-world variable constraints. The task has the static obstacles so that persistence will be effective but will be stale when faced with dynamic environments in the real case.

  3. The parameters are different among the task. For example, how we choose the renoising steps is not clear from the theory and the author only conduct the ablations but not explain the reason behind it. For new tasks, how we select the parameters should be considered.

  4. To my understanding, the method needs training data that includes the feasible solution to the problem. Additionally, if all sampled trajectories violates the constraints then the algorithm will still follow the solution with the least violations. These make the model highly dependent on the pretrained diffusion models. However, safediffuser and projected diffusion can handle additional constraints.

Questions for authors
  1. Could you clarify your contribution to diffusion models or add some experiments like moving obstacles, equality constraints, or real-world applications with visual input or visual constraints?

  2. Could you please check how robust the best parameter selected for a specific task is for different task or introduce how we should select the parameter for general tasks, such as renoising steps?

  3. How does the method perform on the randomly generated constraints when we are not sure if the training data cover the feasible solutions to the problem?

5 comments

We thank the reviewer for the constructive questions. Q3 led to the random-obstacle study in part 2. We have updated the manuscript, and the changes are marked in blue.


W1, Q1: contribution and task difficulty

We clarified the contribution and added two harder task families.

We agree that population warm starting and partial denoising are established. P3 combines selection during denoising with a weighted population retained across control steps. We formalize this process and analyze, under idealized assumptions, when retaining a discovered rare route is easier than repeatedly rediscovering it. The new warm-start comparisons show that reuse accounts for substantial gains, and P3 adds to them, 0.85 against 0.74 for a warm-started Diffusion-ES on Narrow and 0.62 against 0.38 on Turn (Table 1).

On the submitted tasks, the reviewer is right that Avoiding and PointMaze constrain planar positions. AntMaze is an 8-DoF quadruped with a 29-dimensional state and eight torques. On the constraints, P3 only evaluates them on planned trajectories and needs no gradient. They are supplied only at test time, so the pretrained model is never trained against them.

We added Walker2d locomotion under a torso-height ceiling and KUKA sequential block placement with a protection constraint on inactive blocks, the latter a 7-DoF arm with a four-stage objective. On Walker2d, P3 is above the roof on 3.1% of steps against 9.3% for SMC, at less than half the denoiser evaluations. On KUKA, P3 violates the protection constraint in 30% of episodes against 60% for the next best method and places 1.40 of four blocks against 1.20. Both are in §4.4.

Table 21 in Appendix L compares these operations across prior methods, and Table R1 separates the two parts of P3.

Table R1: success without constraint violation on Table 1's eight configurations. The population is redrawn or kept at each control step, and plans are selected at the end of denoising or during it. Bold marks the best value in each column among the rows shown.

MethodPopulationSelectionBothClutt.LeftRightBroadNarrowBlockedTurn
Best-of-NNredrawnat end0.920.340.880.850.740.510.010.43
SMCredrawnduring0.970.480.910.970.750.510.000.53
Persistent best-of-NNkeptat end0.930.870.860.920.750.560.200.37
P3 (ours)keptduring0.990.990.940.980.890.850.210.62

Neither part alone is enough. On Narrow, keeping the population alone reaches 0.56 and selecting during denoising alone 0.51, against 0.85 for P3. Resetting P3's weights at every control step lowers success on the large maze from 0.94 to 0.78 (Table 17).


W2, Q1: changing constraints

We added a study of constraints that change mid-episode, and P3 stays above SMC under every change.

We added Appendix F, with an obstacle that appears on the current plan, an obstacle that moves at two onset times, and a separate study on PointMaze.

Table R2: success without constraint violation under changing constraints on Avoiding. Bold marks the best value in each column.

MethodStillMoves from step 24Moves from step 40Appears on the plan
Projection0.310.330.360.33
SMC0.500.540.470.67
P3 (ours)0.990.960.740.78
P3, no fresh plans0.900.900.890.67

P3 stays above SMC in all four columns, by 0.11 to 0.49. P3 loses 0.03 when the obstacle starts moving at step 24, and 0.25 when the motion starts at step 40, where the variant without fresh plans scores 0.89 against 0.74 for P3. On the PointMaze study, P3 reaches 0.94 against 0.69 for SMC. The revised Limitations state that with a fixed number of fresh plans per step P3 can lose success when the change comes late.


Q1: equality constraints

Equalities enter as absolute residuals, and the Avoiding runs already carry them.

P3 treats an equality like any other constraint. The violation score of eq. (1) takes "the positive violation of an inequality or the absolute residual of an equality" (L171 - 172), and the weighting, re-noising and selection steps do not depend on which kind a constraint is. Every Avoiding run of Table 1 already includes 14 equality constraints, which tie each desired position to the previous one plus the action (L1056 - 1057). The projection baseline runs with 28 of them.

A sampled plan never meets an equality exactly, so we measure how small its residual stays instead of counting feasible plans. Proposition 1 allows a tolerance for this reason.

Visual input is outside our scope, which is receding-horizon control with a state-space trajectory model.

W3, Q2: choosing the re-noising depth

We measured how deep a plan survives, and the depths we use stay within that limit.

The theory says what sets the depth. Corollary 1 keeps a plan on its route when the distance between routes is large relative to the plan spread and the noise added at the depth, so a task with widely separated routes tolerates a deeper re-noise than one with closely spaced routes. The corollary assumes exact denoising and two Gaussian routes, so we measure the limit directly.

We measured this limit on all five tasks. For each task we draw 4,096 plans at fixed states and record the deepest level at which one re-noise and denoise step still keeps half of a plan's main direction of variation. The depths we use are at or below the top of the measured range on every task. The medium maze measures 4 to 5 and uses 4, the large maze measures 14 to 16 and uses 16, Avoiding measures 13 to 16 and uses 10, AntMaze measures 37 to 41 and uses 40, and Walker2d measures 11 to 17 and uses 8. The depth that suits one task therefore does not transfer to another, which is why we measure it for each.

Table R3: sensitivity on Avoiding Cluttered at N=16N=16, out of K=20K=20 reverse steps. Iter. counts reverse diffusion iterations per warm replanning call. Bold marks the best value in each column.

SettingSuccess without violationViolationSwitchesIter.
Cold SMC0.350.03211.4220
t0=2t_0=20.840.1203.183
t0=4t_0=40.980.0113.825
t0=6t_0=60.980.0003.807
t0=8t_0=80.990.0004.129
t0=10t_0=100.990.0044.4611
t0=12t_0=120.970.0004.8613
t0=16t_0=160.820.0025.6217
t0=19t_0=190.400.03111.3620

Success without violation stays within 0.02 across t0=4t_0=4 to 1212, which is a wide plateau rather than a sharp optimum. Below the plateau, a re-noise that is too shallow leaves violations in place, as at t0=2t_0=2. Our rule for a new task is to measure this limit and to re-noise no deeper than it. The measurement above takes one batch of samples at a few states and needs no closed-loop runs.

On the other settings, P3 with 8 particles reaches 0.92 on Clutt. against 0.46 for SMC with 32 (Table 11), so a small population is enough there. Fresh plans are used on Avoiding only, and without them P3 reaches 0.90 on Clutt. against 0.99 with them (Table 6).


W4, Q3: training data coverage

We ran randomly generated obstacle sets, and P3 has the highest success on both families, 0.87 on sparse layouts and 0.54 on dense ones.

The reviewer is right that any sampler that leaves the model's denoising steps unchanged, including best-of-NN, SMC and P3, can only return plans the model supports, and that P3 executes the least-violating plan when every plan violates. Keeping the population makes that case rarer. In 30 logged Clutt. episodes, cold SMC has no acceptable plan at some step in 27 and succeeds in 8, while P3 succeeds in 28 (Table 3 in Appendix A.5). We tested the dependence on training data with 7,000 episodes over randomly placed obstacle sets on Avoiding (Appendix E.5), in a sparse family of three discs and a dense family of six, with positions drawn at random and layouts checked for reachability.

Table R4: success without constraint violation on random layouts, 100 layouts per family across five models. Bold marks the best value in each column.

MethodSparseDense
Unguided0.340.06
SafeDiffuser0.540.24
Projection0.650.30
Best-of-NN0.810.43
SMC0.790.49
Diffusion-ES (warm)0.810.49
P3 (ours)0.870.54

P3 beats every method on the sparse family with all bootstrap intervals above zero, and on the dense family it beats best-of-NN, Projection and SafeDiffuser but ties SMC and warm-started Diffusion-ES. On the 20 layouts that no training demonstration avoids, the samplers reach 0.21 to 0.26 with P3 highest, Projection reaches 0.19 and the unguided model reaches 0.00. No method succeeds without violation on 3 of 100 dense layouts.

On the point about SafeDiffuser and projected diffusion, both are in Table R4. On these constraints, which no model was trained with, they score below P3, 0.30 and 0.24 on the dense family against 0.54.

We are happy to run further checks during the discussion period.

Thanks for the detailed response and the new experiments. They address some of my concerns.

For the new experiments, could the authors explain why they use t0 = 0 in the "Moving obstacle in the large maze" experiment (since the rebuttal reports a range for t0)?

Besides this, I still feel that the paper is more of an engineering work. For the new experiments, it is not clear to me why the proposed algorithm performs better on these tasks, or why these tasks are good benchmark for this proposed algorithm. The new baselines are also not compared on the new tasks.

Reviewer_2sXoOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes P3 as an SMC method for constrained diffusion-based sampling. P3 maintains a weighted population of candidate trajectories across replanning steps, using trajectory shifting, partial re-noising, and constraint-aware weighting and resampling without retraining the diffusion model. The manuscript provides a Feynman–Kac interpretation and some theoretical results. Experiments on limited, simulation-only tasks demonstrate improvements in constraint satisfaction and planning efficiency compared to selected baselines.

Critical strengths
  • Theoretical analysis is provided. The Feynman–Kac formulation provides a mathematical interpretation of accumulated particle weights and Theorem 1 characterizes how re-noising depth affects route retention.
  • The coverage of baselines is decent in experiments.
  • The reported computation efficiency is better than some of the projected or optimization-based baselines.
Critical weaknesses
  1. Writing needs substantial improvements. It is clearly AI-assisted writing, filled with non-standard jargons and “fancy” words (those not quite common in scientific/engineering academic papers) that pose difficulty for readers to quickly understand the underlying idea.
  2. Key novelty is uncleared compared to existing work such as Diffusion-ES [1] and Feynman–Kac diffusion steering [2]. Direct comparisons against such approaches are missing. As such, the specific benefits of persistent SMC was not assessed. The motivation is also not so clear; the paper does not answer the question of "Why does it have to be P3 under which settings?"
  3. Experiments are all low-dimensional navigation tasks, thus proof-of-concept level. More experiments on locomotion and manipulation tasks (in terms of robot planning and control) are necessary to demonstrate the methods applicability across tasks and its practicality. Tasks in other domains such as realistic-resolution image or video generation, molecule or material structure generation would also be cogent.
  4. The method’s performance is not consistent across existing tested tasks: some benchmarks show limited gains or unfavorable tradeoffs in task completion, e.g., both Large and Medium mazes in Table 1 and Maze Medium in Table 3.

References

[1] Yang et al. Diffusion-ES: Gradient-free planning with diffusion for autonomous and instruction-guided driving. CVPR 2024.
[2] Singhal et al. A General Framework for Inference-time Scaling and Steering of Diffusion Models. ICML 2025.

Questions for authors

See weaknesses.

4 comments

We thank the reviewer for the detailed weaknesses, and in particular for the question "Why does it have to be P3 under which settings?", which made us state the operating regime explicitly. We have updated the manuscript, and the changes are marked in blue.


W1: writing

We clarified the terminology and shortened difficult sentences.

The criticism is fair. We made targeted language edits throughout the main text. We replaced clean success with success without constraint violation, rejuvenation with adding fresh plans, and repair with the re-noise and denoise step. We now describe lineage as particle ancestry and define it as the chain of plans a particle descends from. We added short explanations of particles, potentials and resampling. We split long sentences in Sections 3.3 and 3.4 and stated the fixed-observation and exact-denoising assumptions directly. The appendices, tables and figure labels use the same terminology.


W2: novelty and operating regime

We added the missing comparisons, and we now state when P3 helps and when it does not.

We agree that population warm starting and partial denoising are established. P3 combines selection during denoising with a weighted population retained across control steps. We formalize this process and analyze, under idealized assumptions, when retaining a discovered rare route is easier than repeatedly rediscovering it. The new warm-start comparisons show that reuse accounts for substantial gains, and P3 adds to them where the feasible route is rare and long, 0.85 against 0.74 for a warm-started Diffusion-ES on Narrow and 0.62 against 0.38 on Turn.

We ported FK steering and Diffusion-ES from their official code and added both to Table 1, together with TDP and a warm-started Diffusion-ES of our own construction. Diffusion-ES suggests warm starting in one sentence of its Sec. 3.2, but its released planner restarts from noise at every replanning call, so no published warm variant exists and we built it ourselves. Table 21 in Appendix L lists which of these operations each prior method uses.

Table R1: success without constraint violation on Table 1's eight configurations. Bold marks the best value in each column among the rows shown.

MethodBothClutt.LeftRightBroadNarrowBlockedTurn
SMC0.970.480.910.970.750.510.000.53
FK steering0.930.470.880.890.790.420.010.44
Diffusion-ES0.980.380.960.900.760.510.000.43
TDP0.940.380.900.890.760.500.010.42
Diffusion-ES (warm)0.970.980.990.980.860.740.210.38
P3 (ours)0.990.990.940.980.890.850.210.62

On Clutt., Narrow and Blocked, the three methods that redraw their population at every replanning call sit at the level of cold SMC, and the two that keep a population across calls are well above them. That contrast is what isolates persistence, and it was missing from the submission.

On the motivation, the regime is simple to state. P3 helps when a feasible route is rare under the model and has to be held across many replanning calls. Its gain is limited in two cases, and the table shows both. When cold sampling already finds feasible plans, every method is near the ceiling, as on Both and Left. When the model almost never produces a feasible plan, no sampler does well, as on Blocked, where keeping the population raises success from 0.00 to 0.21 and the best of all thirteen methods reaches 0.24.

The Discussion (§5) of the revision now states when persistence helps and when it adds little.

W3: task scope

We added a locomotion task and a manipulation task, and P3 has the lowest violation rate on both.2

First, on the submitted tasks. The reviewer is right that Avoiding and PointMaze constrain planar positions. AntMaze is an 8-DoF quadruped, with a 29-dimensional state and eight torques driven by an inverse-dynamics head, so it is locomotion rather than navigation.

Second, we added two task families. We added Walker2d locomotion under a 1.3 m torso-height ceiling, using the pretrained Diffuser walker2d-medium-expert model, and KUKA sequential block placement with a protection constraint on the three inactive blocks, a 7-DoF arm with a four-stage objective. Neither model was trained with its constraint. Both are in §4.4 and Table 2 of the revision, with details in Appendices J and K.

Table R2: Walker2d, 48 episodes per method, roof at 1.3 m. Steps violating is the share of walked steps above the roof. Overshoot is the mean excess above the roof over walked steps. NFE counts particle denoiser evaluations per control step. Bold marks the best value in each column.

MethodSteps violating (%)Overshoot (mm)D4RL scoreNFE/step
Unguided16.54.30105.61280
Best-of-NN8.71.30107.01280
SafeDiffuser15.53.56105.91280
SMC9.31.45106.91280
FK steering6.80.89107.01280
P3 (ours)3.10.20106.8577

P3 is above the roof on 3.1% of steps against 6.8% for the next best method, with a mean overshoot of 0.20 mm against 0.89 mm, and uses less than half the denoiser evaluations of the cold methods while matching the unconstrained return. On KUKA, P3 violates the protection constraint in 30% of episodes against 60% for the next best method and places 1.40 of four blocks against 1.20.

On image, video, molecule and material generation, those settings have no receding horizon, so there is no population to carry between calls. P3 targets receding-horizon control.


W4: inconsistent gains

On these configurations every selection-based method is limited by what the model samples, and P3 leads or ties that group.

P3, SMC and the other selection-based methods leave the model's denoising steps unchanged and only select among its samples, so they can return only plans the model produces. Table 1 shows where this limits them.

On the large maze, one 64-step plan already spans the obstacles, so correcting that plan by constrained optimization works well (§5). DiRecT on both configurations and Projection on Narrow are therefore ahead of every selection-based method. Among the selection-based methods P3 is the best on both configurations, and on Narrow it plans about five times faster than DiRecT and far faster than Projection (Table 10).

On the medium maze, the model rarely samples the required detour, so there is little to select from. The selection-based methods that redraw their plans at every step succeed in at most 1% of episodes. The methods that carry plans across steps reach the highest rates there, 0.20 to 0.24, and P3 is one of them (§4.2).

On Table 3 (Table 5 in the revision), task success counts reaching the goal irrespective of violations. On Blocked, SMC reaches the goal as often as the unguided model and never without a violation, so its higher task success comes from paths that violate the constraint. P3 is the selection-based method that most often reaches the goal without a violation, tied with one other. Our claim is about success without constraint violation.

We are glad to run further comparisons during the discussion period.

Thank you for the detailed rebuttal and additional experiments.

I appreciate the authors' efforts in expanding the task scope, which largely addresses my concern W3.

However, my main concerns regarding W2 and W4 remain:

  • W2: The additional baselines are helpful, but the results suggest that much of P3's improvement comes from trajectory warm-starting, which is an established technique. Warm-started Diffusion-ES already achieves comparable or even better performance on Clutt., Left, Right, and Blocked. As such, I remain unconvinced, through extensive empirical studies, that the proposed method establishes a sufficiently substantial algorithmic contribution or unique advantages beyond existing approaches.
  • W4: The empirical limitations remain: optimization-based methods outperform P3 on the large mazes, while success without constraint violations remains low on the blocked medium maze. The rebuttal explains these outcomes but does not fully resolve the concern about the method's effectiveness across different settings.

Overall, the rebuttal strengthens the paper by adding the new tasks, but my primary reservations regarding algorithmic novelty and performance consistency still remain. Therefore, I would maintain my original rating.

We thank the reviewer for the reply and are glad that the new tasks largely address W3. We answer W2 and W4 below. The PDF is updated again, and the changes are marked in blue.


W2: contribution beyond warm starting

The warm-started Diffusion-ES is not an existing approach. We built it for this comparison, and where warm starting alone gives no gain, P3 still does.

We agree that warm starting is an established idea and that it accounts for a substantial part of the gain. None of the released population baselines uses it, however. FK steering, TDP and the released Diffusion-ES redraw their plans at every replanning call, and earlier diffusion planners warm-start one trajectory (Table 21). The warm-started row is our construction (Appendix B), so the four configurations the reviewer lists compare two methods that keep a population, both from this paper. On Left and Right every population method is already between 0.85 and 0.99. On Clutt. and Blocked, Diffusion-ES as released reaches 0.38 and 0.00, and the warm start we added takes it to 0.98 and 0.21.

Table R3: where keeping the population gives the gain, and where P3 adds to it. SMC and Diffusion-ES as released redraw their plans at every call. Bold marks the best value in each row, and both values where the two methods that keep their plans tie.

TestSMCDiffusion-ES (released)Diffusion-ES (warm, ours)P3 (ours)
Clutt.0.480.380.980.99
Blocked0.000.000.210.21
Narrow0.510.510.740.85
Turn0.530.430.380.62
Obstacle appears on the plan0.670.630.670.78
KUKA, episodes with a violation70%65%50%30%
Walker2d, steps above roof9.3%6.5%3.3%3.1%
Maze, disc slides0.690.650.990.94

On Clutt., Blocked, Walker2d and the sliding-disc maze, keeping the population gives the gain, and the warm control is ahead on the maze. On Turn and when an obstacle appears on the plan, the warm control gives nothing over SMC, and on Turn it is below the released Diffusion-ES. Only P3, which keeps its plans and selects during denoising, is above all three (with the obstacle on the plan, +0.11+0.11 over the warm control, 95% CI [0.04,0.17][0.04, 0.17]). On Narrow and KUKA, P3 adds to the gain from keeping. The paper also explains when keeping pays, with one Feynman–Kac model over denoising and replanning, Theorem 1 on re-noising depth and Proposition 1 on kept plans against fresh draws.

The general comment reports FK steering, Diffusion-ES and TDP on the new tasks. Our earlier reply quoted the next method at 6.8% on Walker2d and 60% on KUKA. With the warm control added, the closest method is at 3.3% and 50%.


W4: consistency across settings

No method in Table 1 is best everywhere, and P3 is the most consistent of the thirteen. It is best or tied in four of the eight configurations and within 0.13 of the best method in all eight.

The reviewer is right on both points, and the paper states them (§4.2 and §5). DiRecT is ahead on both large-maze configurations and Projection on Narrow, and no method exceeds 0.24 on Blocked. Table 1 also shows what those methods give up elsewhere.

Table R4: consistency over the eight configurations of Table 1, computed from its entries. Bold marks the best value.

MethodBest or tiedLargest gap to the best method
DiRecT2 of 80.71 (Clutt.)
Projection0 of 80.53 (Clutt.)
SMC0 of 80.51 (Clutt.)
Diffusion-ES (warm)2 of 80.24 (Narrow)
P3 (ours)4 of 80.13 (Narrow)

DiRecT wins the large maze and falls to 0.28 on Clutt., where P3 reaches 0.99. On Blocked, DiRecT and Projection reach 0.12 and 0.11, about half of P3's 0.21, so correcting one plan by optimization does not resolve that configuration either. The pretrained model rarely samples the required detour, and every method that redraws its plans stays at or below 0.01. P3 also needs only constraint evaluations, with no gradient and no solver, and plans about five times faster than DiRecT on Narrow (Table 10).

Comments on the paper

3 comments

We thank the reviewers for their comments. Three concerns are shared across the reviews, and we answer them together here. The reply to each reviewer gives the details. We have updated the manuscript, and the changes are marked in blue.


2sXo W2, qzZx W1, MNKC W1: contribution

We clarified the contribution and added four baselines to Table 1, and P3 matches or beats all four on seven of eight configurations.

We agree that population warm starting and partial denoising are established. P3 combines selection during denoising with a weighted population retained across control steps. Table 1 now includes FK steering, TDP, Diffusion-ES and a warm-started Diffusion-ES that we built and matched to P3's compute. Reuse accounts for substantial gains, and P3 adds to them where the feasible route is rare and long, 0.85 against 0.74 for the warm-started baseline on Narrow and 0.62 against 0.38 on Turn. Resetting P3's carried weights at every control step lowers success on the large maze from 0.94 to 0.78 (Table 17). Table 21 in Appendix L lists which operations each prior method uses.


2sXo W3, qzZx W1, MNKC W2: task scope

We added two task families beyond planar obstacle avoidance, and P3 violates the constraint least on both.

On Walker2d locomotion under a torso-height ceiling, P3 is above the roof on 3.1% of steps against 6.8% for the next best method, at less than half the denoiser evaluations. On KUKA sequential block placement with a protection constraint, P3 violates the constraint in 30% of episodes against 60% for the next best method, and places 1.40 of four blocks against 1.20. Neither model was trained with its constraint. Both tasks are in §4.4, Table 2 and Appendices J and K.


qzZx W2, MNKC Q2: changing constraints

We tested obstacles that move or appear without warning, and P3 stays ahead of the baselines in every test.

In the new Appendix F, an obstacle appears on the current plan or moves across the workspace. In the three tests with a change, P3 stays above SMC by 0.11 to 0.42 (Table 16). On the large maze with a disc that slides across a corridor, P3 reaches 0.94 against 0.69 for SMC (Table 17).


Update: the new baselines on the new tasks

We ran FK steering, Diffusion-ES, its warm-started variant and TDP on Walker2d, KUKA and the changing-constraint tests. P3 stays ahead of the three that redraw their plans. The warm-started variant is ahead of P3 in only one test, the sliding-disc maze.

Reviewer qzZx asked for these rows, and they are in Tables 2, 16, 17, 19 and 20. FK steering, Diffusion-ES and TDP redraw their plans at every step, as SMC does. On Walker2d they leave 4.1% to 6.8% of steps above the roof against 3.1% for P3, with more than twice the denoiser evaluations. TDP goes down to 0.1% if its roof gradient is made three times stronger, but the walker then falls in 15 of 48 episodes. On KUKA they have a violation in 65% to 75% of episodes against 30% for P3. In the four tests where the constraint changes, all three stay at or below SMC, and P3 is above SMC by 0.11 to 0.42.

The warm-started Diffusion-ES is our own construction and keeps its plans between steps, as P3 does. It is close to P3 on Walker2d, at 3.3% against 3.1%. P3 is ahead on KUKA, with a violation in 30% of episodes against 50%, and when an obstacle appears on the plan, at 0.78 against 0.67 (paired difference 0.11, 95% CI [0.04,0.17][0.04, 0.17]). The warm-started variant is ahead on the maze with a sliding disc, at 0.99 against 0.94. On that maze, resetting P3's carried weights at every step lowers its success from 0.94 to 0.78.

The "next best method" figures quoted above predate these rows. They are now 3.3% on Walker2d and 50% on KUKA.

We thank the Area Chair and the reviewers for the discussion. The results below are in the revised PDF, marked in blue.


Contribution

P3 selects during denoising on a weighted population of diffusion plans that it keeps across replanning calls. None of the cited methods does this, and the paper explains when it pays.

  • The paper contributes a formulation and an analysis. Denoising and replanning form one Feynman–Kac particle model. Under exact denoising at a fixed observation, Theorem 1 gives how re-noising depth affects whether a kept plan stays on its route. Proposition 1 gives the probability that a population holds no acceptable plan, for kept plans and for fresh draws.
  • We measured that quantity on a model we did not train. On Walker2d, a population that is redrawn at every step ends with no plan under the roof at 26% of control steps for SMC, against 10% for the population P3 keeps (Appendix J).
  • No released population baseline keeps its plans. FK steering, TDP and the released Diffusion-ES redraw them at every replanning call, and earlier diffusion planners warm-start one trajectory (Table 21). The warm-started Diffusion-ES in our tables exists because we built it at the request of MNKC.
  • P3 adds to warm starting. Against that control, P3 reaches 0.85 against 0.74 on Narrow and 0.62 against 0.38 on Turn, where warm starting alone falls below cold SMC. It has a violation in 30% of KUKA episodes against 50%, and reaches 0.78 against 0.67 when an obstacle appears on the plan (+0.11+0.11, 95% CI [0.04,0.17][0.04, 0.17]). Among the new tests, the control is ahead of P3 only on the sliding-disc maze.

What the discussion added

The revision adds the baselines, task families and changing-constraint tests the reviewers asked for. P3 beats every baseline that redraws its plans on both new tasks and in every changing-constraint test.

  • Baselines (2sXo, MNKC): FK steering, Diffusion-ES, TDP and the warm-started Diffusion-ES are in Table 1 and on every new task. P3 matches or beats all four on seven of Table 1's eight configurations.
  • Tasks beyond planar avoidance: Walker2d under a height ceiling and KUKA block placement with a protection constraint (§4.4, Table 2, Appendices J and K). P3 has the lowest violation rate on KUKA, and on Walker2d among the methods that never fall. On Walker2d it uses less than half the denoiser evaluations of the baselines that redraw.
  • Consistency (2sXo): no method in Table 1 is best everywhere. P3 is best or tied in four of its eight configurations and within 0.13 of the best method in all eight. DiRecT, which wins the large maze, is 0.71 behind on Clutt..
  • TDP (MNKC): same model, scenes, cost, grading and denoiser budget. P3 places 1.40 of four blocks against 0.70 on the 20 scenes of Table 20, and 1.30 against 0.83 over all 30 scenes.
  • Changing constraints (qzZx, MNKC): obstacles that move or appear on the plan (Appendix F). In the four tests with a change, P3 is above the best baseline that redraws its plans by 0.11 to 0.42.
  • Re-noising depth (qzZx): one measurement on the pretrained model bounds the depth for a new task. Sweeps on Avoiding and on the moving-disc maze show that success is flat over a wide range below the bound.
  • Training-data coverage (qzZx): 7,000 episodes on randomly placed obstacles, where P3 has the highest success on both families (Appendix E.5).
  • Writing (2sXo, MNKC): we replaced the non-standard terms, added short definitions and split long sentences.

The comparisons the reviewers asked for support the paper's claim. Keeping the population across replanning calls carries most of the gain, and selection during denoising with carried weights adds to it where feasible plans are rare.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.