{"id":"54b749cc-b31d-49cd-8497-5970b530c444","arxiv_id":"2509.08160","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A MAPF-inspired search guided by single-arm and dual-arm diffusion models plans collision-free motions for many arms without higher-order training data.","lead":"DG-MAP plans motions for many robot arms by training separate diffusion models for single arms and for pairs of arms, then using search to resolve collisions. It achieves high success rates with 3 to 8 arms while requiring only single- and dual-arm training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pairwise repair can cycle on three-arm deadlocks; the 'collisions are primarily pairwise' assumption is not stress-tested.","rationale":"The paper's central contribution is the claim that a MAPF-style decomposition using only single- and dual-arm diffusion models can achieve high multi-arm success without higher-order training data. The linchpin is the 'collisions are primarily pairwise' assumption. If that assumption fails—for instance, in scenarios where a three-arm coordinated maneuver is necessary—the method's performance could degrade sharply, because the search can indefinitely cycle through pairwise repairs. The authors are transparent about this in Appendix A.1, but they do not test the failure regime. My proposed adversarial benchmark directly probes this. I do not think the concern invalidates the paper: the empirical results are strong, the limitation is disclosed, and the approach is a reasonable engineering heuristic. However, the claim is not yet fully established for general multi-arm coordination, which is exactly the conditional verdict. I therefore leave the reader's verdict unchanged. The missing error bars and unequal compute comparisons are secondary; the pairwise assumption is the most load-bearing.","tokens_in":11366,"tokens_out":8980,"duration_ms":113671,"concrete_test":"Generate 100 adversarial scenarios with three 6-DoF arms that must pass through a single narrow gate simultaneously, with start/goal assignments arranged so that any pairwise avoidance creates a collision with the third arm (a three-arm rotation/swap). Run DG-MAP with the same hyperparameters as Table 4 and record success rate, fraction of episodes ending in 'best effort' timeout, and number of Repair calls. If success falls below the 88.8% observed for 8-arm hard tasks (or below a pre-registered threshold such as 50%), the pairwise assumption is the limiting factor. A complementary check is to compare against a variant that trains a three-arm diffusion model; if the gap is large, higher-order coordination is essential.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central scaling claim rests on an untested assumption: that resolving collisions one pair at a time always suffices for N>2 arms. In the search, Repair (Alg. 3) always samples a brand-new plan for the ego arm and never adds a constraint on the third arm; a plan that avoids arm j may hit arm k, and a later repair for that new conflict can recreate the original collision with j. Such cyclic three-arm deadlocks are precisely the 'complex coordination strategies involving three or more arms simultaneously' that Appendix A.1 admits may not be captured. Because the benchmark defines difficulty by pairwise workspace overlap (Fig. 3) and the timeout returns a 'best effort' plan (Alg. 1, line 26), the reported >88-98% success does not demonstrate that the pairwise decomposition resolves genuinely three-way coordination demands. The method may still work well on typical random scenes, but the central claim 'efficiently scales to larger number of arms' is only as strong as this pairwise-decomposition assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DG-MAP, a multi-arm motion planner that combines two conditional diffusion models—one trained on single-arm BiRRT trajectories and one on dual-arm interactions—with a MAPF-inspired search. At each receding-horizon step, each arm proposes B candidate delta-action sequences from the single-arm model; the search detects the earliest pairwise collision and generates successors by either re-selecting an existing candidate (Rebranch) or sampling a new candidate from the dual-arm model conditioned on the conflicting partner (Repair). The planner executes up to the earliest predicted collision time and re-plans. Experiments in PyBullet with 3–8 six-DoF arms on goal-reaching tasks and a four-arm pick-and-place task compare DG-MAP against the learning-based baselines of Ha et al. (2020), trained on limited or extended interaction data. The paper reports success rates above 88–98% for DG-MAP, outperforming the limited-data baseline and slightly outperforming or matching the extended-data baseline, while using only single- and dual-arm training data.","tokens_in":11647,"tokens_out":5528,"duration_ms":67838,"significance":"If the claims hold, DG-MAP is a useful step toward data-efficient multi-arm planning: it avoids collecting large N-arm demonstration datasets and leverages a structured search to resolve collisions that a purely end-to-end policy misses. The combination of diffusion-based trajectory generation with MAPF-style conflict resolution is sensible, and the pick-and-place result (89% success vs. 71.4% for the extended-data baseline) is encouraging. The paper does not ship machine-checked proofs or code, and its central scaling claim rests on an explicit pairwise-decomposition assumption that is not stress-tested. The empirical evaluation is also missing variance and compute normalization, so the strength of the current evidence is moderate rather than decisive.","major_comments":[{"comment":"The pairwise-repair mechanism is load-bearing for the N>2 scaling claim, but the paper does not show that the search terminates or avoids three-arm cycles. Repair(i,j) samples a fresh plan for i conditioned only on j; it imposes no constraint that the new plan avoid a third arm k, and a later repair of i against k can recreate a collision with j. A concrete deadlock is possible: plans a1,b1,c1 collide between a1 and b1; Repair yields a2 avoiding b1 but colliding with c1; Repair of a2 against c1 yields a3 avoiding c1 but colliding again with b1. The bookkeeping set K records only tried plan indices, not pairwise spatio-temporal constraints, so the same conflict can reappear. Appendix A.1 concedes that 'complex coordination strategies involving three or more arms simultaneously might not be fully captured by this pairwise approach,' and the benchmark difficulty is defined by pairwise works","section":null},{"comment":"All success rates are point estimates with no error bars, confidence intervals, trial counts, or multiple seeds. This is particularly problematic for the small differences claimed against Baseline-ED (e.g., N=8 easy 0.951 vs 0.951; N=3 easy 0.980 vs 0.984). Without variance information, the statements 'consistently outperforms' and 'remains competitive' are not statistically supported. Please report the number of scenarios per cell, seeds, and mean±std or CIs; the pick-and-place result should also include variance across trials.","section":null},{"comment":"The comparison is not compute-normalized. DG-MAP is a search-based planner with a 60 s planning timeout and B=10 candidate samples per arm, while the baselines are end-to-end learned policies. The success-rate tables do not reflect planning latency, and Figure 5 shows only a coarse runtime comparison without error bars or timing methodology. Since the central claim is 'efficiently scales to larger number of arms,' the paper should report per-query wall-clock time, success rate as a function of planning timeout, and a compute-normalized comparison (e.g., same total CPU/GPU budget for all methods).","section":null},{"comment":"The planner cannot recover if all initial single-arm proposals contain intrinsic collisions; Appendix A.1 acknowledges this but does not quantify it. The dual-arm model only handles inter-arm conflicts, so the overall success rate is upper-bounded by the single-arm model's ability to produce at least one feasible trajectory per arm. Please report how often failure is caused by the single-arm proposals versus by inter-arm search failure, and consider an ablation that injects known-good single-arm trajectories to separate the two effects.","section":null}],"minor_comments":[{"comment":"The variable t* is used at line 18 but its initialization is not specified. Please initialize t* to T_p or define it explicitly.","section":null},{"comment":"The notation for the conflict bookkeeping set K is ambiguous. The text says 'conflicting plan indices' but the set is used per arm in Algorithms 2 and 3. Define K as a set of (arm, plan-index) pairs or as per-arm sets, and update node construction accordingly.","section":null},{"comment":"The loss expressions contain unmatched brackets and nested norms that are hard to read. Please clean up the notation, e.g., write E[||ϵ - ϵ_θ(...)||²].","section":null},{"comment":"The phrase 'consistently high average success rates above 90%' is contradicted by the hard-difficulty cells for N=6 and N=8 (0.888 in both). Please clarify whether the claim refers to the 'Average' column only or to all settings, and quote the correct range (e.g., >88%).","section":null},{"comment":"The paper would benefit from releasing code and trained diffusion models; the project website currently points to a teaser page but not to a repository. This is not required for acceptance but would substantially improve reproducibility.","section":null}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the reported results are promising, but the pairwise-decomposition assumption is the central load-bearing element and it is only weakly validated. The missing error bars and compute-normalized comparisons make it hard to assess whether the improvement over Baseline-ED is real. I would support a major revision that adds targeted experiments (three-arm deadlocks, repeated seeds, runtime curves) rather than rejecting outright, because the framework is defensible and the limitations are acknowledged in the appendix."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this is a solid engineering paper that does what it claims on simulation, but the central pairwise-decomposition assumption is honestly flagged and not stress-tested, so take the 'efficiently scales to larger number of arms' claim as demonstrated for pairwise-dominated scenes.\n\nThe new combination is real: single-arm and dual-arm conditional diffusion models inside a MAPF-style conflict search, where the dual-arm model is conditioned on the conflicting arm's mirrored observations to generate repairs. That is not in the cited prior work. The experiments are reasonably convincing: DG-MAP beats a limited-data MARL baseline by a large margin and edges out an extended-data baseline, on goal-reaching from 3 to 8 arms and on a pick-and-place task. The Appendix's DiffusionQL comparison is a nice robustness check; the planner seems to be the main driver, not the specific diffusion training objective.\n\nThe main soft spot is the one the authors themselves concede in A.1: pairwise repair may not capture three-way coordination. The stress-test concern is real: Repair (Alg. 3) always resamples a plan for the ego arm conditioned only on one other arm, and never blocks the third arm, so cyclic deadlocks are possible in principle. The paper does not report any experiment designed to force such three-way conflicts; difficulty is defined by pairwise workspace overlap. So the >90% success rates should be read as 'works well when pairwise conflicts dominate,' which is a useful but more limited claim.\n\nOther softer spots: no error bars or variance reported for any success rate; the comparison to baselines is not compute-normalized (DG-MAP performs search with up to 60s timeout); and no code or data are released, which makes the results hard to reproduce. These are not fatal, but they matter for a methods paper.\n\nWho should read this: anyone working on multi-arm manipulation or learned motion planning with search. It is a credible demonstration that a small set of pairwise diffusion models plus a simple search can handle teams of up to 8 arms in simulation. It deserves a serious referee, and the revision should address the three-way stress test, add error bars, and release code/data.","headline":"Pairwise-diffusion + MAPF search is a real and useful combination, but the central scaling claim needs a stress test for three-arm coordination and error bars.","tokens_in":12061,"tokens_out":2312,"would_cite":true,"duration_ms":27254,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DG-MAP claims pairwise diffusion models plus search plans up to eight robot arms with over 90% success, using only single- and dual-arm training data.","keywords":["multi-arm motion planning","diffusion models","Multi-Agent Path Finding","conflict resolution","pairwise decomposition","scalable planning","receding horizon","manipulation"],"falsifier":"A concrete test is a three-arm task in which all three arms must pass through a narrow shared bottleneck simultaneously with no way to sequentialize the passage: if DG-MAP's pairwise repair cycles fail or success rates drop well below the reported 90%, the pairwise-decomposition assumption is the cause.","tokens_in":11304,"feed_emoji":"🤖","tokens_out":2643,"duration_ms":32188,"temperature":0.7,"pith_summary":"This paper proposes a multi-arm motion planner that combines two conditional diffusion models with a Multi-Agent Path Finding (MAPF)-style search. One model generates feasible single-arm trajectories; a second generates collision-avoiding trajectories for a pair of arms. The search interleaves rebranching among pre-sampled plans and repairing conflicts with the dual-arm model. The central claim is that this pairwise decomposition scales to teams of 3–8 arms while staying above roughly 90% success, even though the models are trained only on single- and dual-arm demonstrations. This matters because it suggests that costly higher-order multi-arm training data may be unnecessary for many coordinated manipulation tasks.","feed_headline":"Two-arm training coordinates eight arms at 90% success","feed_subtitle":"A diffusion-guided search over pairwise conflict repairs plans 3–8 robot arms without higher-order training data.","key_machinery":"Two conditional denoising diffusion models serve as proposal generators: the single-arm model (epsilon_theta1) samples future joint-action sequences conditioned on an arm's own observation history, and the dual-arm model (epsilon_theta2) samples ego-arm actions conditioned on a paired observation that concatenates the conflicting arm's transformed state with the ego-arm's state. A best-first search over plan-index tuples, using a cost function of smoothness, goal proximity, and collision penalty, explores alternatives through Rebranch (trying other pre-sampled single-arm plans) and Repair (regenerating plans with the dual-arm model). The collision-free duration t* is used for receding-horizo","core_discovery":"DG-MAP shows that a MAPF-inspired decomposition—plan each arm independently, detect the earliest pairwise collision, and repair that pair with a dedicated dual-arm diffusion model—can produce collision-free plans for up to eight 6-DoF arms. In experiments, the planner exceeds 90% average success across easy, medium, and hard goal-reaching tasks and all team sizes, outperforms a learning-based baseline trained on the same limited single/dual-arm data, and stays competitive with a baseline trained on richer three- and four-arm data. It also achieves 89% success on a four-arm pick-and-place task. The paper's core claim is that pairwise interaction modeling, combined with search over candidate t","pith_inferences":["A natural stress test is a task where three or more arms must simultaneously pass through a narrow shared bottleneck; if DG-MAP fails or times out there, the pairwise assumption itself would be the limiting factor.","The dual-arm diffusion model could be reused as a collision-avoidance prior in trajectory optimization or as a local planner inside other hierarchical multi-robot systems, not just within this specific search.","When pairwise repair repeatedly fails, an adaptive mechanism that switches to a higher-order interaction model—trained on triple-arm data when available—could extend the framework to intrinsically three-arm-coordinated scenarios.","The same single/dual-arm decomposition may transfer to heterogeneous robot teams, provided the paired observation transformation is redefined for different morphologies,"],"forward_implications":["If the central claim holds, multi-arm planners can be trained from single- and dual-arm demonstrations, removing the need for expensive full-team data collection.","The search-over-diffusion-proposals architecture converts a generative model into a repair mechanism, so a collision encountered during search directly conditions the next sampling round.","Receding-horizon execution up to the predicted collision-free time allows the planner to adapt to dynamic conditions while keeping computation bounded.","The reported pick-and-place success (89%) indicates the approach generalizes beyond simple goal-reaching to multi-stage manipulation tasks.","The pairwise-decomposition principle, if correct, suggests that most practical multi-arm conflicts are local interactions between two arms, so scalable planning can focus on resolving those local conflicts."],"supporting_citations":[{"why":"Supplies the baseline learning-based multi-arm planner, the expert-demonstration setup for single- and dual-arm BiRRT data, the task-difficulty metric, and the pick-and-place evaluation environment that DG-MAP compares against and builds on.","marker":"[14]"},{"why":"Provides the Diffusion Policy framework (CNN UNet with FiLM conditioning) used as the backbone for both the single-arm and dual-arm conditional diffusion models.","marker":"[23]"},{"why":"Defines Multi-Agent Path Finding and its decomposition into single-agent planning plus conflict resolution, the conceptual template for DG-MAP's planning structure.","marker":"[13]"},{"why":"Supplies Conflict-Based Search, the MAPF algorithm whose idea of resolving pairwise conflicts by imposing constraints on replanning directly inspires DG-MAP's Rebranch and Repair successor generation.","marker":"[22]"},{"why":"Establishes Denoising Diffusion Probabilistic Models, the base generative formulation whose noise-prediction objective and denoising loop both diffusion models are trained with.","marker":"[16]"},{"why":"Represents a prior approach integrating single-robot diffusion models with MAPF for multi-robot navigation, which DG-MAP extends to articulated manipulators with a dedicated dual-arm diffusion model.","marker":"[20]"}],"fun_headline_variants":["Pairwise diffusion scales multi-arm planning to eight arms","Eight arms, one pairwise diffusion model: 90% success","No higher-order data needed: pairwise diffusion plans eight arms","Diffusion on pairs plans eight arms at 90% success"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Multi-arm collisions can be resolved one pair at a time: a diffusion model trained only on two interacting arms, together with search, suffices to coordinate any number of arms, so coordination that intrinsically requires three or more arms moving at once is not modeled.","fun_headline_variants_meta":{"raw":{"variants":["Pairwise diffusion scales multi-arm planning to eight arms","Eight arms, one pairwise diffusion model: 90% success","No higher-order data needed: pairwise diffusion plans eight arms","Diffusion on pairs plans eight arms at 90% success"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000762,"raw_usage":{"total_tokens":3195,"prompt_tokens":694,"completion_tokens":2501,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":2442}},"tokens_in":438,"tokens_out":2501,"duration_ms":19782,"temperature":1.0,"reasoning_tokens":2442,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:07:56.966701+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test is a three-arm task in which all three arms must pass through a narrow shared bottleneck simultaneously with no way to sequentialize the passage: if DG-MAP's pairwise repair cycles fail or success rates drop well below the reported 90%, the pairwise-decomposition assumption is the cause.","supporting_citations":[],"review_version":1}