{"id":"2874fb2b-b054-46a6-a16a-68a418eeee54","arxiv_id":"2508.19199","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A diffusion model trained on optimized resolution labels picks which regions of a deformable object need high-resolution dynamics for a planning query, yielding 2.1x faster planning with a small cost increase.","lead":"This paper learns to automatically simplify a robot's physics model of a deformable tree, region by region, so planning runs about twice as fast while still moving the right branch. It uses a diffusion model that predicts which tree regions need high detail for each start-to-goal manipulation query.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Eq. 4 'dynamics-accuracy' objective is actually a task-cost against the observed final state, so the two-stage labels may be biased and the learned generator may be learning suboptimal resolutions.","rationale":"The paper's stated contribution is that query-guided, task-specific resolution selection can cut planning cost with little task-performance loss. What has to be true is that the ω* labels used to train the diffusion generator are reasonable optima of a task-performance/complexity tradeoff. Eq. 4 is the only 'fast' stage that makes the two-stage pipeline practical; if it does not actually measure dynamics accuracy, then the second stage, which is also under-specified because Algorithm 2's termination test appears to continue when all regions are simplified and because ε_tol and δ are not reported, starts from a prior with no demonstrated relation to the objective. Figure 7 could still be a valid black-box result even with poor labels, but the central claim would no longer establish the proposed mechanism. I agree with the reader's weakest assumption, and the concrete test directly replaces the suspect objective with a real prediction error and re-runs the headline experiment. The reader's CONDITIONAL verdict is appropriate; this concern does not by itself falsify the empirical result but it prevents unconditional acceptance.","tokens_in":11966,"tokens_out":6900,"duration_ms":68222,"concrete_test":"Re-run the first stage with a true dynamics-accuracy objective, e.g. minimize ||decode(f̂_ω(s1,a1:T)) − sT|| + wdyn|ω|1, regenerate the ω* labels through Algorithm 2, retrain the diffusion generator, and re-run the Section V-C evaluation on the 100 held-out queries; if the 2.1x speedup or the cost gap versus full changes materially, the reported result depends on the mis-specified Eq. 4.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section V-C rests on the quality of the resolution labels ω* used to train the diffusion generator. That quality depends on the chained optimization in Section IV-D. The first stage is described as a dynamics-accuracy prior, but Eq. 4 minimizes cplan(BUILDGRAPH(sT,ω), zG) + wdyn|ω|1. BUILDGRAPH(sT,ω) is not a rollout; it is a graph encoding of the observed final state at resolution ω, so the first term measures how well that simplified encoding of the final state scores against the goal in the planning cost, not how accurately the simplified model predicts the trajectory. The sentence immediately after Eq. 4, saying that BUILDGRAPH(sT,ω) 'computes a rollout,' contradicts the formula. If this term does not correlate with true predictive error or with closed-loop task cost, the initialization fed to Algorithm 2 is biased, and the greedy simplification in Algorithm 2, whose termination condition 'while ω* unchanged ≥ Igrace iterations or Σω*_t = 0' is ambiguous or misprinted, may stop at labels that are not the intended task-performance optima. The learned generator would then learn a distribution over suboptimal resolutions, and the experimental comparison in Figure 7 would not support the proposed mechanism even if the headline numbers are reproducible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for automatically generating spatially adaptive dynamics models for deformable-object manipulation. Given a planning query, defined by start and goal point clouds, a diffusion-based model generator outputs per-region resolution parameters for a graph neural network dynamics model used in MPPI planning. To obtain training labels, the authors design a two-stage optimization that first minimizes a dynamics-accuracy cost (Eq. 4) and then refines the resolution vector under a closed-loop task-performance constraint (Algorithm 2). The method is evaluated on a simulated tree-manipulation task, where it achieves a 2.1x reduction in planning time over a full-resolution model with a reported cost increase of 0.006. The paper also analyzes the distribution of predicted resolutions and compares against full, minimal, and mode baselines.","tokens_in":12344,"tokens_out":9524,"duration_ms":81931,"significance":"The central idea is timely and relevant: rather than hand-specifying model complexity, the method learns task-specific resolutions from planning-query data, with a practical two-stage procedure that limits the number of costly closed-loop evaluations. The reported speedup on a non-trivial tree-manipulation task is a concrete falsifiable result, and the use of a diffusion model to capture multimodal resolution distributions is appropriate. The paper explicitly provides architecture details and MPPI hyperparameters, which aid reproducibility. However, the validity of the central claim depends heavily on the correctness of the label-generation pipeline, and several presentation and specification issues currently weaken confidence in that pipeline.","major_comments":[{"comment":"The text states that BUILDGRAPH(sT,ω) \"computes a rollout,\" but BUILDGRAPH is defined in Section IV-A as the function that encodes a state into a graph at resolution ω; it does not perform a rollout. As written, Eq. (4) minimizes cplan(BUILDGRAPH(sT,ω), zG) + wdyn|ω|1, which is the planning cost between the simplified encoding of the observed final state and the goal graph plus a sparsity penalty, not a prediction error. If the implementation follows Eq. (4), the first stage is not a dynamics-accuracy prior and the initialization for Algorithm 2 may be biased; if the implementation actually performs a rollout, the equation and the text must be corrected. Please clarify the intended objective and, if necessary, re-evaluate the impact on the learned resolutions.","section":"IV-D.1, Eq. (4)"},{"comment":"The termination condition in line 5, \"while ω* unchanged ≥ Igrace iterations or Σω^t = 0 do,\" is ambiguous and appears to contain a misprint. It is unclear whether the loop continues while either condition holds or stops when either holds, and the expression \"Σω^t\" is not defined. This condition determines when the resolution-label search terminates and therefore directly affects the quality of the training data; please restate it as a precise Boolean expression.","section":"IV-D.2, Algorithm 2"},{"comment":"The dataset size is reported inconsistently: Section IV-B says \"We train on a dataset of 6075 samples holding out 10% for validation,\" the paragraph after Algorithm 2 says \"dataset of 11K s1,sG,ω* tuples,\" and Section V-A says \"5500 (s1,sG,ω*) tuples were in the dataset.\" These numbers must be reconciled, since the training-set size is important for assessing generalization claims.","section":"IV-B, IV-D.2, V-A"}],"minor_comments":[{"comment":"The sentence \"This section outlines our approach for generating a resolution ω* which in order to construct a dataset mapping planning queries as (s1,sG) pairs to an optimized ω\" is grammatically malformed and should be reworded.","section":"IV-D, first paragraph"},{"comment":"The text references Figure 7, but the actual box plot is not present in the manuscript text I reviewed; the figure should be included to support the reported means and standard deviations.","section":"V-C"},{"comment":"The sentence \"The complexity of the graph, and thus the computation time for computing scales quadratically with the number of vertices and edges in z_t\" is incomplete; clarify the intended statement.","section":"IV-A"},{"comment":"The chamfer-distance cost cplan has unclear notation: the bounds \"i=1^P\" and \"j'<P\" are not consistent, and the mask m is used without a clear definition in this context.","section":"IV-C"},{"comment":"Please specify how the \"mode\" baseline is computed: whether it is the single most frequent full resolution vector in the training set or per-segment modes.","section":"V-C"},{"comment":"The phrase \"The optimization process (Algorithm. 2)\" contains a typo; \"Algorithm. 2\" should be \"Algorithm 2.\"","section":"IV-D.2"}],"recommendation":"major_revision","confidential_remarks":"I believe the paper fits the journal's scope. The main risk is internal consistency of the data-generation pipeline; if the authors clarify the Eq. 4 objective and fix Algorithm 2, the contribution is publishable. I do not consider the self-referential nature of the label optimization a fatal flaw, but the authors should discuss the potential domain gap when the same simulator and planner are used for both label generation and evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Core take: the paper has a real idea, and the headline result is believable. It learns to map a planning query (start/goal pointclouds) to per-region binary resolutions for a GNN dynamics model, so MPPI plans on a cheaper graph without sacrificing closed-loop cost. The diffusion model is a sensible choice for multimodal label distributions. The more distinctive contribution is the two-stage label-generation scheme: cheap dynamics-based prior first, then expensive closed-loop refinement. I don't know of prior work doing query-conditioned spatial resolution selection; the closest cited work optimizes a global resolution or geometric simplification, so the novelty claim holds.\n\nThe empirical comparison is credible: 2.1x planning speedup and a 0.006 cost gap versus full resolution, with mode and minimal baselines worse, on held-out queries. Figure 6 shows the learned distribution tracks the optimized labels across query classes. The authors are also honest that the speedup is modest because they downsample from 1317 to 227 points.\n\nThe main soft spot is that the stress-test note about Eq. 4 mostly lands. The surrounding text says BUILDGRAPH(sT, omega) 'computes a rollout' and that cplan measures prediction error, but the equation is cplan(BUILDGRAPH(sT, omega), zG) + wdyn|omega|_1, which is a task-cost in the simplified graph space against the goal, not a prediction error. That is a misdescription, and it matters because the stage-two closed-loop optimization starts from this prior. A biased prior is not fatal because the second stage directly optimizes closed-loop cost, but the greedy bit-flipping in Algorithm 2 has an ambiguous termination condition ('while omega* unchanged >= Igrace iterations or sum omega*_t = 0') and the tolerance epsilon_tol is never reported, so I cannot tell whether the labels are near the intended optimum. Add to that three different dataset sizes (6075, 11K, 5500) and no code or data, and the reproducibility bar is not met yet.\n\nThe circularity worry is weaker than it looks. Yes, the labels come from the same simulator, dynamics model, and MPPI planner used in evaluation, but the evaluation is on held-out queries and the point is to show query-conditioned selection beats query-independent baselines. That is standard in sim. What is missing is an oracle baseline that runs MPC with the optimized omega* labels on test queries; without it, you can't tell how much the learned generator loses relative to the labels.\n\nOverall, this is a paper for people working on deformable-object MPC or adaptive model simplification. The problem framing and the two-stage label idea are worth citing. It deserves a serious referee. I would send it to review with a request to fix the Eq. 4 description, report all hyperparameters, resolve the dataset-size inconsistency, and add the oracle-omega* planning baseline.","headline":"A genuinely useful query-conditioned resolution-selection idea with a plausible speedup on a tree task, but the label-optimization details are too inconsistent to fully trust the mechanism yet.","tokens_in":12746,"tokens_out":3908,"would_cite":true,"duration_ms":35995,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A generative model learns query-specific model resolutions, doubling planning speed for deformable-object manipulation.","keywords":["deformable object manipulation","model-based planning","graph neural networks","diffusion model","adaptive model resolution","model predictive control","tree manipulation"],"falsifier":"On a held-out set of planning queries, compute the correlation between the stage-1 objective (Eq. 4) and the closed-loop task cost achieved by MPC with the corresponding resolution vector; if the correlation is near zero or negative, the prior is not carrying the argument. Alternatively, run Algorithm 2 from a random initialization instead of the stage-1 prior and compare final resolutions and closed-loop costs: if random initialization finds substantially better or cheaper solutions, the prior is not a reliable starting point.","tokens_in":11723,"feed_emoji":"🌳","tokens_out":7281,"duration_ms":61764,"temperature":0.7,"pith_summary":"This paper tries to establish that a learned, query-conditioned model generator can replace a fixed full-resolution dynamics model for model-based deformable-object manipulation, preserving closed-loop task performance while sharply reducing planning time. The central proposal is to predict, from the start and goal pointclouds of a planning query, which spatial regions of a deformable object should be modeled at high resolution and which can be collapsed to a single node in a graph neural network dynamics model. To train this predictor without running expensive closed-loop rollouts for every candidate resolution, the authors construct a dataset through a two-stage optimization: first minimize a dynamics-accuracy objective in the simplified graph space, then refine the resolution vector against actual MPC closed-loop cost within a tolerance. On a simulated tree-manipulation task, the method achieves a 2.1x speedup in planning time over full-resolution models at a small, quantified cost increase. If the approach generalizes, it offers a path from planning data to task-specific, computationally efficient models for new manipulation tasks.","feed_headline":"Query-guided model cuts planning time 2.1x, keeps task quality","feed_subtitle":"Learned per-region resolutions trim model-based planning cost with little loss on a simulated tree manipulation task.","key_machinery":"The load-bearing machinery is a diffusion-based model generator $p_\\theta(\\omega \\mid z_{1-G})$ trained on a dataset of planning queries paired with optimized binary resolution vectors. The query is encoded as a single graph $z_{1-G}$ built from full-resolution encodings of start and goal states, with edges connecting corresponding particles; the diffusion model denoises binary resolution vectors conditioned on this graph. The other half of the machinery is the two-stage dataset-construction optimization: Eq. 4 initializes $\\omega$ by minimizing the plan cost in the simplified graph space against the observed final state plus an $\\ell^1$ penalty on high-resolution regions, and Algorithm 2 refines $\\omega$ by running actual MPC and accepting simplifications whose closed-loop task cost stays within $\\epsilon_{\\mathrm{tol}}$ of the best cost seen. The GNN dynamics model itself, with complexity scaling quadratically in graph size, is what makes resolution selection consequential.","core_discovery":"The central claim is that region-specific model resolution can be learned as a function of the planning query: given a start state, a goal state, and a segmentation of the object pointcloud, a conditional diffusion model outputs a binary resolution vector $\\omega$ that determines which segments are represented at high resolution and which are collapsed to a single vertex. The dataset for this mapping is generated by a chained optimization in which a dynamics-accuracy-constrained simplification (Eq. 4) produces an initialization, and a closed-loop task-performance-constrained optimization (Eq. 5) then simplifies further while keeping MPC cost within $\\epsilon_{\\mathrm{tol}}$ of the best observed cost. The experiments on a tree-manipulation task report that the learned generator matches the distribution of optimized resolutions across query classes and yields a 2.1x planning-time speedup over full-resolution models with a 0.006 increase in task cost, corresponding to roughly 1 cm extra average distance for moved particles.","pith_inferences":["The same query-conditioned resolution idea could be applied to other high-dimensional model classes, such as particle-based fluid or cloth simulators, by treating the segmentation as an additional input rather than a given.","A natural extension, flagged by the paper as future work, is to learn the regions themselves from raw pointclouds instead of relying on a provided segmentation; that would make the generator applicable to new object shapes without manual region definitions.","The closed-loop refinement stage could be made more sample-efficient by reusing the planner's own trajectory data across queries, potentially replacing the per-query MPC evaluations with a shared value model.","An untested implication is that the learned resolution distribution should transfer to unseen tree geometries if the GNN dynamics model generalizes across graph structures; if transfer fails, the pointcloud conditioning alone may be insufficient."],"forward_implications":["If the method is correct, a model-based planner can use a query-specific simplified graph instead of a full-resolution graph, reducing planning time without retraining the planner or the dynamics model.","The two-stage optimization offers a recipe for building training data for task-informed model simplification without paying full closed-loop MPC cost for every candidate resolution.","The observed couplings between segments (for example, the top of the tree often staying high-resolution when a nearby branch moves) imply that the learned resolutions capture message-passing dependencies, not just geometric motion.","Because the reported speedup grows with graph size, applying the approach to larger pointclouds than the 1317-particle tree should yield larger planning-time savings at similar performance tolerance.","The fact that a query-independent mode baseline is both slower and worse than the query-guided generator supports the conclusion that conditioning on the start and goal states is doing the work."],"supporting_citations":[{"why":"Supplies the denoising diffusion training and sampling procedure used for the model generator.","marker":"[34]"},{"why":"Supplies the graph-neural-network dynamics model that resolution vectors parameterize.","marker":"[3]"},{"why":"Supplies the model predictive path integral planner used for closed-loop evaluation.","marker":"[39]"},{"why":"Supplies the simulated deformable-object environment for query generation and task-cost evaluation.","marker":"[37]"},{"why":"Supports the premise that simplified, low-resolution dynamics models can still be effective for control.","marker":"[9]"},{"why":"Provides the dynamic-resolution baseline that the paper's spatially local resolution approach extends.","marker":"[18]"}],"fun_headline_variants":["Query-guided models double planning speed, keep task performance","Adaptive model resolution yields 2.1x faster planning","Diffusion model picks where to detail deformable objects","Task-aware resolution: 2.1x speedup in deformable planning","Learn model detail from query, cut planning time in half"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The first optimization stage assumes that a cost computed in the simplified graph space against the observed final state is a reliable proxy for how accurately the simplified model predicts motion; if that proxy is misleading, the learned resolution labels start from a biased initialization and may be suboptimal.","fun_headline_variants_meta":{"raw":{"variants":["Query-guided models double planning speed, keep task performance","Adaptive model resolution yields 2.1x faster planning","Diffusion model picks where to detail deformable objects","Task-aware resolution: 2.1x speedup in deformable planning","Learn model detail from query, cut planning time in half"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000246,"raw_usage":{"total_tokens":1520,"prompt_tokens":908,"completion_tokens":612,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":528}},"tokens_in":524,"tokens_out":612,"duration_ms":6089,"temperature":1.0,"reasoning_tokens":528,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:53:40.430762+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out set of planning queries, compute the correlation between the stage-1 objective (Eq. 4) and the closed-loop task cost achieved by MPC with the corresponding resolution vector; if the correlation is near zero or negative, the prior is not carrying the argument. Alternatively, run Algorithm 2 from a random initialization instead of the stage-1 prior and compare final resolutions and closed-loop costs: if random initialization finds substantially better or cheaper solutions, the prior is not a reliable starting point.","supporting_citations":[{"cited_title":"Aggressive driving with model predictive path integral control,","cited_arxiv_id":null,"evidence_quote":"Supplies the model predictive path integral planner used for closed-loop evaluation."},{"cited_title":"Softgym: Benchmarking deep reinforcement learning for deformable object manipulation,","cited_arxiv_id":null,"evidence_quote":"Supplies the simulated deformable-object environment for query generation and task-cost evaluation."},{"cited_title":"The surprising effectiveness of linear models for visual foresight in object pile manipulation,","cited_arxiv_id":null,"evidence_quote":"Supports the premise that simplified, low-resolution dynamics models can still be effective for control."}],"review_version":2}