{"id":"f8f16ea7-a81a-445d-97da-7d60c53e2cf2","arxiv_id":"2607.06499","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":5,"one_line_summary":"CE-MPPI resolves MPPI averaging-induced failure by pruning colliding rollouts, clustering feasible ones via DBSCAN on a geometric direction feature, and selecting a single cluster for the control update.","lead":"This paper modifies MPPI, a sampling-based robot motion planner, by pruning colliding rollouts, clustering the survivors by direction, and selecting one cluster for the control update. It addresses a real failure mode where averaging incompatible trajectories causes robots to hesitate or collide near obstacles.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Central claim rests on 3 runs with no variance and a 179s baseline reflecting catastrophic MPPI failure rather than typical operation; the λ parameter is confounded with the method comparison and no sensitivity analysis is provided.","rationale":"The reader's CONDITIONAL verdict is appropriate. The reader correctly identified evaluation thinness and the Eq. 18 generalization concern. I agree on the evaluation thinness but would emphasize a more specific load-bearing issue: the confounding of the λ parameter with the method comparison, combined with the absence of variance estimates and the extreme baseline operating point (179s). These are more fundamental than the Eq. 18 heuristic concern because they affect whether the core claim — that the clustering architecture itself resolves the failure — is supported at all, regardless of dynamic obstacle scenarios.\n\nThe algorithmic idea is sound in principle: pruning colliding rollouts, clustering by a geometric direction feature, and selecting a single cluster for the update is a reasonable approach to avoiding averaging across incompatible modes. The geometric feature (Eq. 12–13) is a reasonable design choice for separating left/right passing modes. The within-cluster renormalization (Eq. 19–20) is correct. There is no internal inconsistency in the formulation.\n\nHowever, the paper has 5 hand-tuned parameters (K, N, λ, DBSCAN ε, dt) with no sensitivity analysis, no code release, and only 3 runs per condition. The claim of \"architectural resolution\" is strong relative to the evidence base. The verdict should remain CONDITIONAL pending broader evaluation with statistical reporting and parameter sensitivity analysis.","tokens_in":11300,"tokens_out":3374,"duration_ms":234647,"concrete_test":"Re-run the real-world UR5e experiment (or an equivalent high-fidelity simulation) with at least 10 trials per method, reporting mean ± std. Additionally, sweep λ over {0.01, 0.05, 0.1, 0.3, 0.7} for both MPPI and CE-MPPI on the same scenario. If the 48% improvement shrinks below 20% at matched λ, or if MPPI's time-to-goal varies by more than 2× across λ values, the headline claim that CE-MPPI architecturally resolves the failure mode (rather than benefiting from parameter selection) would be significantly weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper claims CE-MPPI \"architecturally resolves\" averaging-induced failure, but the experimental evidence is insufficient to establish this at the claimed level. Three specific issues:\n\n(1) All quantitative results are averaged over only 3 runs with no variance or standard deviation reported (Tables II and IV). With n=3, the confidence interval on any mean is extremely wide, and the 17.4% improvement over CSC-MPPI in Scenario 2 (9.66s vs 7.98s) could easily be within noise.\n\n(2) The headline 48% time-to-goal reduction comes from a real-world experiment where standard MPPI took 179.20s to traverse 0.83m — an average end-effector speed of ~0.005 m/s, indicating near-complete stagnation. This is an extreme failure mode, not a representative operating point. Any method that breaks the deadlock would show dramatic relative improvement, inflating the headline number.\n\n(3) The temperature parameter λ differs dramatically across settings: 0.7 for Scenario 1, 0.01 for Scenario 2, 0.05 for real-world. Since λ directly controls the sharpness of importance weighting — and thus whether averaging-induced failure occurs at all — the choice of λ is confounded with the method comparison. A fairer test would tune λ for the baseline or report results across a λ sweep. Without this, it is unclear whether CE-MPPI's advantage comes from the clustering architecture or from the baseline operating at a particularly unfavorable λ.\n\nThe reader's concern about Eq. 18 (dynamic obstacle heuristic) is valid but secondary: the more load-bearing issue is that the entire experimental base is too thin and too confounded to support the breadth of the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper proposes Clustering-Embedded MPPI (CE-MPPI), which augments standard MPPI with three stages: (1) pruning colliding rollouts, (2) DBSCAN clustering of feasible rollouts using a geometric direction feature derived from collision-derived reference points, and (3) obstacle-aware cluster selection (minimum cost for static obstacles, opposite-to-obstacle-motion for dynamic obstacles). The within-cluster MPPI update then uses only rollouts from the selected cluster. The method is evaluated in two 2-D simulation scenarios (static and dynamic obstacles) against standard MPPI and CSC-MPPI, and on a 6-DoF UR5e manipulator against standard MPPI. The authors report improvements in time-to-goal and path length in all experiments.","tokens_in":11760,"tokens_out":1390,"duration_ms":332228,"significance":"The paper addresses a well-known practical limitation of MPPI—averaging-induced failure in multi-modal environments—and proposes a constructive, modular solution. The geometric direction feature (Eqs. 12–13) anchored to collision-derived reference points is a sensible improvement over velocity-based clustering features used in prior work (CSC-MPPI), and the obstacle-aware selection criterion for dynamic scenes (Eq. 18) addresses a gap in the existing literature. The real-world deployment on a UR5e with GPU-accelerated rollouts demonstrates practical feasibility. However, the experimental evaluation has significant limitations (small sample sizes, confounded parameters, limited scenario diversity) that constrain the strength of the claims that can be drawn from the evidence.","major_comments":[{"comment":"Tables II and IV: All quantitative results are averaged over only 3 runs with no variance or standard deviation reported. With n=3, confidence intervals on the means are extremely wide. For example, the 17.4% improvement over CSC-MPPI in Scenario 2 (9.66s vs. 7.98s) could easily fall within noise. The authors should report standard deviations and, ideally, increase the number of runs to support the comparative claims.","section":null},{"comment":"Table I and Table III: The temperature parameter λ differs dramatically across settings (0.7 for Scenario 1, 0.01 for Scenario 2, 0.05 for real-world). Since λ directly controls the sharpness of importance weighting—and thus whether averaging-induced failure occurs at all—the choice of λ is confounded with the method comparison. A fairer test would tune λ for the baseline or report results across a λ sweep. Without this, it is unclear whether CE-MPPI's advantage comes from the clustering architecture or from the baseline operating at a particularly unfavorable λ.","section":null},{"comment":"Table IV: The headline 48% time-to-goal reduction comes from a real-world experiment where standard MPPI took 179.20s to traverse 0.83m—an average end-effector speed of approximately 0.005 m/s, indicating near-complete stagnation. This represents an extreme failure mode rather than a representative operating point. Any method that breaks the deadlock would show dramatic relative improvement, inflating the headline number. The authors should either provide a more representative baseline or explicitly contextualize this as a worst-case comparison.","section":null},{"comment":"Eq. 18 and Section III-B.2: The dot-product heuristic for dynamic obstacle selection—choosing the cluster whose average motion is most opposite to the obstacle's motion direction—is validated on a single dynamic scenario with one obstacle moving along the x-axis. In environments with multiple moving obstacles or obstacles whose motion is not aligned with the robot's nominal direction, this criterion could select a cluster that is safe from one obstacle but collides with another, or that is suboptimal overall. The paper should discuss this limitation and, at minimum, test with a second dynamic-obstacle configuration (e.g., obstacle moving along y-axis, or two obstacles).","section":null}],"minor_comments":[{"comment":"Section III-C.1: The statement that CSC-MPPI 'exceeds the 10 Hz control budget' on the UR5e is stated without quantitative timing data. Reporting the per-iteration computation time for each method would strengthen this claim.","section":null},{"comment":"Algorithm 1, line 7: The fallback to standard MPPI when |C_col_t| = 0 is mentioned, but the frequency of this fallback in practice is not reported. This information would help readers understand how often the clustering pipeline is actually invoked.","section":null},{"comment":"Eq. 13: The small constant ε is introduced for numerical stability but its value is not specified in the simulation or real-world parameter tables.","section":null},{"comment":"Section III-A.2: The robot model is described as a differential-drive platform, but the heading angle θ_t has no cost weight in Q (Q = diag(10, 10, 0)). This should be clarified—whether the heading is irrelevant to the task or whether this is a deliberate design choice.","section":null},{"comment":"Fig. 5: The figure caption references sub-panels (a)–(k) but the distinction between time-steps within each method's row is not clearly explained in the caption. A brief note on what each sub-panel represents (e.g., different time-steps) would improve readability.","section":null},{"comment":"Section III-D.2: The collision penalty w_c = 1000 is described as 'sufficiently large' but no analysis of its sensitivity or interaction with the clustering pipeline is provided. A brief discussion would be helpful.","section":null},{"comment":"Table I: The DBSCAN ε parameter is listed only for the real-world experiment (Table III, 0.3 m) but not for the 2-D simulation. The simulation DBSCAN parameters should be reported.","section":null}],"recommendation":"major_revision","confidential_remarks":"The paper presents a reasonable and practically motivated algorithm, but the experimental evaluation is not yet at the standard expected for the field. The n=3 sample size with no variance reporting is below the norm, and the λ confound is a genuine threat to the central comparative claim. The 179s baseline is particularly concerning as it likely reflects a degenerate operating point. I would encourage the authors to address these issues, as the core algorithmic ideas are sound and the paper could be a solid contribution with stronger evaluation. The dynamic obstacle heuristic (Eq. 18) is the weakest conceptual point—it is a hand-designed heuristic with limited validation—but this is a scope concern rather than a correctness flaw."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. The referee correctly identifies that our experimental evaluation has limitations in sample size, parameter confounding, baseline representativeness, and scenario diversity for the dynamic-obstacle heuristic. We agree with all four major comments and will revise the manuscript accordingly. Specifically, we will (1) increase the number of runs and report standard deviations, (2) conduct a λ-sweep for the baseline and report results, (3) contextualize the real-world headline number as a worst-case comparison and add a more representative baseline, and (4) add a second dynamic-obstacle configuration and discuss the limitations of the dot-product heuristic for multi-obstacle settings. One practical constraint—re-running the real-world UR5e experiments with additional configurations—requires hardware access, but we are committed to completing this before resubmission.","responses":[{"response":"The referee is correct. With n=3 and no reported variance, the comparative claims are not adequately supported. We will increase the number of runs for both simulation scenarios (to at least 10) and for the real-world experiment (to at least 5), and report means and standard deviations in all results tables. If any comparison loses statistical significance after this, we will adjust the corresponding claims accordingly.","revision_made":"yes","referee_comment":"Tables II and IV: All quantitative results are averaged over only 3 runs with no variance or standard deviation reported. With n=3, confidence intervals on the means are extremely wide. The 17.4% improvement over CSC-MPPI in Scenario 2 could easily fall within noise. The authors should report standard deviations and, ideally, increase the number of runs."},{"response":"This is a valid and important concern. The different λ values across scenarios were chosen based on per-scenario tuning, but we did not control for whether the baseline was operating at a favorable or unfavorable λ. We will conduct a λ-sweep for both standard MPPI and CSC-MPPI in each scenario and report performance across multiple λ values. This will allow readers to assess whether CE-MPPI's advantage is robust to the baseline's λ setting or whether part of the gain is attributable to parameter confounding. We will present these results either as a table or a figure in the revised manuscript.","revision_made":"yes","referee_comment":"Table I and Table III: The temperature parameter λ differs dramatically across settings (0.7 for Scenario 1, 0.01 for Scenario 2, 0.05 for real-world). Since λ directly controls the sharpness of importance weighting—and thus whether averaging-induced failure occurs at all—the choice of λ is confounded with the method comparison. A fairer test would tune λ for the baseline or report results across a λ sweep."},{"response":"We agree with this assessment. The standard MPPI baseline in the real-world experiment was in a near-stagnation regime, and the 48% figure is inflated by this extreme failure mode. We will take two corrective steps: (1) We will explicitly contextualize the real-world result as a comparison against a worst-case MPPI failure mode, rather than a representative operating point, and adjust the abstract and conclusions to avoid implying that 48% is a representative improvement. (2) We will add an additional real-world baseline condition where MPPI operates at a more favorable λ or with a higher collision penalty that reduces stagnation, providing a more representative comparison point. We acknowledge that this requires additional hardware experiments, which we will conduct before resubmission.","revision_made":"yes","referee_comment":"Table IV: The headline 48% time-to-goal reduction comes from a real-world experiment where standard MPPI took 179.20s to traverse 0.83m—an average end-effector speed of approximately 0.005 m/s, indicating near-complete stagnation. This represents an extreme failure mode rather than a representative operating point. Any method that breaks the deadlock would show dramatic relative improvement, inflating the headline number. The authors should either provide a more representative baseline or explicitly contextualize this as a worst-case comparison."},{"response":"The referee's concern is well-founded. The dot-product criterion in Eq. 18 was designed for and tested in a single-obstacle scenario, and its behavior in multi-obstacle or non-axis-aligned settings is not validated. We will address this in two ways: (1) We will add a second dynamic-obstacle simulation in which the obstacle moves along the y-axis (perpendicular to the robot's nominal direction), and report comparative results. If feasible within the revision timeline, we will also test a two-obstacle configuration. (2) We will add an explicit discussion of the limitation of the current heuristic for multi-obstacle environments, noting that the single-dot-product criterion does not account for interactions between multiple obstacle motion directions and may require extension (e.g., a weighted combination over obstacles or a different selection criterion) in such settings. We will frame this as a scope limitation and a direction for future work.","revision_made":"yes","referee_comment":"Eq. 18 and Section III-B.2: The dot-product heuristic for dynamic obstacle selection is validated on a single dynamic scenario with one obstacle moving along the x-axis. In environments with multiple moving obstacles or obstacles whose motion is not aligned with the robot's nominal direction, this criterion could select a cluster that is safe from one obstacle but collides with another, or that is suboptimal overall. The paper should discuss this limitation and, at minimum, test with a second dynamic-obstacle configuration."}],"tokens_in":11273,"tokens_out":1141,"duration_ms":177776,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"CE-MPPI is a reasonable engineering improvement to MPPI that addresses a real failure mode. The core idea — prune colliding rollouts, cluster the survivors using a geometric direction feature anchored to where collisions tend to terminate, then do the weighted update within a single cluster — is clean and well-specified. The geometric feature (Eq. 12–13) is a genuine improvement over CSC-MPPI's velocity-based clustering: using the direction from the collision-derived reference point to each rollout's terminal position is a more natural way to separate left-passing from right-passing modes, and the intuition for why velocity statistics can fail to separate these is correct. The algorithm is clearly written and the pseudocode is reproducible. Credit where it's due: the method is simple, the motivation is sound, and the within-cluster renormalization is the right thing to do once you've committed to a cluster. The dynamic-obstacle selection criterion (Eq. 18, pick the cluster moving most opposite to obstacle motion) is a reasonable heuristic for the coupling problem, though it's tested on exactly one scenario with one obstacle moving along the x-axis. That's thin but the idea itself is not crazy. The soft spots are real and mostly in the experiments. Three runs with no variance reported is not enough to trust the 17.4% improvement over CSC-MPPI in Scenario 2 — that could easily be noise. The headline 48% real-world improvement comes from a baseline that took 179 seconds to traverse 0.83 meters, which is near-complete stagnation rather than a representative operating point. Any method that breaks the deadlock would look dramatic by comparison. The most concerning issue is the temperature parameter: λ is 0.7 in Scenario 1, 0.01 in Scenario 2, and 0.05 in real-world. Since λ directly controls weight sharpness and thus whether averaging-induced failure occurs, the method comparison is confounded. A λ sweep or at least a baseline tuned at the same λ would make the claims much stronger. CSC-MPPI is excluded from real-world experiments due to computational overhead, which is a legitimate practical reason but limits the comparison. The central algorithmic idea holds up and is worth publishing. But the evaluation needs more runs, variance reporting, and a λ sensitivity analysis before the claims can be taken at face value. This paper is for MPPI practitioners working on cluttered manipulation or navigation — they'll find the geometric clustering feature directly useful. It deserves a serious referee who can push for the additional experiments. I'd recommend major revision: the idea is sound, the evidence isn't there yet.","headline":"CE-MPPI: clustering with geometric direction features for MPPI","tokens_in":12102,"tokens_out":1114,"would_cite":false,"duration_ms":81867,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Cluster rollouts by collision direction, cut MPPI hesitation by 48%","keywords":["MPPI","motion planning","DBSCAN clustering","averaging-induced failure","obstacle avoidance","robotics","trajectory optimization","manipulator control"],"falsifier":"Construct a dynamic environment with two obstacles moving in opposite directions along the robot's path. If the dot-product opposition criterion selects a cluster that avoids one obstacle but drives the robot into the other, the heuristic fails and the selection logic would need to be replaced with a multi-obstacle-aware criterion.","tokens_in":11543,"feed_emoji":"🤖","tokens_out":1186,"duration_ms":185042,"temperature":0.7,"pith_summary":"Model Predictive Path Integral (MPPI) control is a sampling-based robot motion planner: it stochastically perturbs a nominal control sequence, forward-simulates hundreds of parallel rollouts, and updates the control via an importance-weighted average of those rollouts. The paper identifies a structural failure mode in this averaging step: when an obstacle blocks the path, some feasible rollouts go left and some go right, but averaging their perturbations produces a control update that points toward the obstacle, causing hesitation or collision. The authors propose Clustering-Embedded MPPI (CE-MPPI), which inserts a pruning and clustering stage between rollout generation and the weighted update. First, all colliding rollouts are pruned. Second, the mean terminal position of those colliding rollouts is computed as a reference point. Third, each surviving rollout is assigned a geometric feature: the normalized direction vector from that collision-derived reference point to the rollout's own terminal position. DBSCAN clusters these direction vectors into groups that correspond to distinct avoidance modes (e.g., left-passing vs. right-passing). Fourth, a single cluster is selected—by minimum average cost for static obstacles, or by maximum directional opposition to obstacle motion for dynamic obstacles—and the MPPI weighted update is computed using only rollouts within that selected cluster. This ensures the averaged perturbations are always from a single, coherent avoidance mode, eliminating the cross-mode cancellation that causes averaging-induced failure. On a UR5e 6-DoF manipulator, CE-MPPI achieves a 48% reduction in time-to-goal and a 12% shorter end-effector path compared to standard MPPI, primarily by eliminating near-obstacle hesitation.","feed_headline":"Cluster rollouts by collision direction, cut MPPI hesitation by 48%","feed_subtitle":"Pruning colliding trajectories and grouping survivors by a geometric direction feature stops MPPI from averaging incompatible avoidance roll","key_machinery":"The collision-derived reference point (mean terminal position of colliding rollouts), the normalized direction feature from this point to each feasible rollout's terminal position, DBSCAN clustering of these features, and the dual selection logic (minimum-cost for static, dot-product opposition for dynamic).","core_discovery":"The central mechanism is the collision-derived reference point: the mean terminal position of all colliding rollouts. By measuring the direction from this reference to each feasible rollout's terminal state, CE-MPPI obtains a geometric feature that naturally separates rollouts into clusters by which side of the obstacle they pass on. This is a sharper clustering signal than the velocity-based statistics used in prior work (CSC-MPPI), because two rollouts passing on opposite sides of an obstacle can have similar average velocities but necessarily have different direction vectors relative to the collision centroid. The within-cluster update then guarantees that the importance-weighted average只","pith_inferences":["The dot-product opposition criterion for dynamic obstacles is validated only on a single obstacle moving along one axis. In environments with multiple moving obstacles, selecting the cluster most opposite to one obstacle's motion could steer the robot toward another obstacle. A multi-obstacle extension would need a vector sum or priority-weighted combination of opposition scores.","The method reverts to standard MPPI when no colliding rollouts are observed (|C_col| = 0), which means the clustering benefit is only activated once the robot has already sampled trajectories that hit obstacles. In sparse-obstacle environments where the robot approaches at high speed, the first time collisions are detected may be too late for clustering to prevent a crash, suggesting a proactive c","The geometric direction feature is defined in task space (planar or end-effector position). For high-DoF manipulators where joint-space configuration matters for self-collision or joint-limit avoidance, clustering purely on end-effector direction may miss infeasible modes that share a terminal direction but differ in joint configuration."],"forward_implications":["The pruning-and-clustering approach could generalize to any sampling-based planner that uses a weighted average over diverse rollouts, including trajectory optimization methods that blend sampled candidates.","The collision-derived reference point concept could extend to higher-dimensional configuration spaces where 'left vs. right' generalizes to distinct homotopy classes of paths around obstacles.","For dynamic obstacles, replacing the hand-designed dot-product heuristic with a learned or model-predictive selection criterion could handle multi-obstacle scenarios where a single opposition direction is insufficient.","The method's reliance on GPU-parallel rollout generation and pruning suggests that as parallel compute scales, the clustering stage's overhead becomes negligible relative to the simulation cost, making the approach increasingly practical."],"fun_headline_variants":["Cluster MPPI rollouts by collision direction to stop hesitation near obstacles","Use collision-derived reference points to split MPPI rollouts into feasible modes","Group MPPI survivors geometrically to avoid averaging incompatible rollouts","CE-MPPI clusters rollouts by which obstacle side they pass on","Direction-from-collision-centroid clusters MPPI rollouts better than velocity stats"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The dot-product heuristic for dynamic obstacles—picking the cluster whose average motion is most opposite to the obstacle's motion direction—is validated on a single scenario with one obstacle moving along the x-axis. Whether this criterion generalizes to multiple moving obstacles or obstacles whose motion is not aligned with the robot's nominal direction is untested.","fun_headline_variants_meta":{"raw":{"variants":["Cluster MPPI rollouts by collision direction to stop hesitation near obstacles","Use collision-derived reference points to split MPPI rollouts into feasible modes","Group MPPI survivors geometrically to avoid averaging incompatible rollouts","CE-MPPI clusters rollouts by which obstacle side they pass on","Direction-from-collision-centroid clusters MPPI rollouts better than velocity stats","Prune colliding rollouts, cluster by direction, cut MPPI time-to-goal 48%","Cluster MPPI rollouts by terminal direction relative to collision mean","Avoid MPPI averaging failure by clustering feasible rollouts by obstacle side","Geometric direction feature separates MPPI rollouts passing opposite obstacle sides","CE-MPPI selects clusters by cost in static scenes, obstacle flux in dynamic ones"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":828,"prompt_tokens":609,"completion_tokens":219,"prompt_tokens_details":null},"tokens_in":609,"tokens_out":219,"duration_ms":19961,"temperature":1.0,"reasoning_tokens":40,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T03:32:49.684452+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Construct a dynamic environment with two obstacles moving in opposite directions along the robot's path. If the dot-product opposition criterion selects a cluster that avoids one obstacle but drives the robot into the other, the heuristic fails and the selection logic would need to be replaced with a multi-obstacle-aware criterion.","supporting_citations":[],"review_version":1}