{"id":"ecb32e53-08fe-4d19-aec0-ac8476e38546","arxiv_id":"2607.09587","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Diffusion students distilled from a privileged MAPPO oracle recover multi-modal cooperative actions under partial observability and beat local MARL and deterministic cloning on three multi-robot tasks.","lead":"CoDiMAD trains a fully informed multi-robot teacher, then distills it into local-only student robots that sample actions with diffusion models instead of averaging them. This matters because communication-free robot teams often face ambiguous local views that hide several valid joint strategies.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged offline-support limitation.","rationale":"The reader's strongest claim accurately tracks the paper's argument and evidence (theory sketch + Table II + action/trajectory visualizations). The weakest assumption the reader isolates is precisely the softest point in the pipeline, and the authors themselves flag it. Because that limitation is already priced into the CONDITIONAL verdict, and because no independent correctness risk (circular evaluation, mis-specified baselines, or broken theory) emerges on a second pass, the verdict does not need to move. The concrete test above simply operationalizes the already-noted offline-support concern so that acceptance can be finalized once code and the unfiltered ablation are available.","tokens_in":13558,"tokens_out":510,"duration_ms":5903,"concrete_test":"Release the promised code and re-train both CoDiMAD and BC-RNN on the identical filtered dataset while also reporting an unfiltered (all-oracle-rollouts) ablation; if the CoDiMAD–BC-RNN gap collapses under the filtered set or if unfiltered data reverses the ranking, the multi-modality advantage is overstated. Otherwise the claim stands as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (multi-modal conditional oracle actions under partial observability; MSE mode-averaging vs. diffusion recovery; Table II gains) is internally consistent with the paper's evidence. Proposition 1 / Corollary 1 correctly diagnose why deterministic BC-RNN can produce invalid intermediate actions, and Fig. 3–4 plus the BC-RNN collision spikes supply direct empirical support. The reader's weakest assumption—Stage-2 filtering to deterministic means of successful, collision-free oracle episodes only (§IV-B, Alg. 1 lines 17–19)—is real and is already acknowledged by the authors in §VI as an offline-distillation limitation. It does not, however, falsify the reported comparisons under the evaluated support: the same filtered dataset is used for both CoDiMAD and BC-RNN, so the large gaps (especially Box Pushing 72.2 % vs. 6.5 % and the ~19× collision reduction on Coverage) still isolate the effect of modeling the conditional distribution rather than its mean. No stronger, previously unstated load-bearing flaw (e.g., hidden communication, reward leakage, or theory–experiment mismatch) is required for the claim to hold within the paper's stated scope.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes CoDiMAD, a three-stage privileged-distillation framework for communication-free multi-robot coordination under partial observability. A MAPPO oracle is trained with global state, an offline dataset of local-observation–oracle-action pairs is collected from successful collision-free rollouts (using the oracle mean action), and decentralized student policies are trained as conditional DDPMs (with DDIM inference) that act only on local BEV observations. The central claim is that partial observability induces multi-modal conditional oracle-action distributions (Proposition 1), that MSE-based deterministic distillation collapses to a low-density conditional mean between modes (Corollary 1), and that diffusion students recover coherent modes, yielding higher task success and fewer collisions than MAPPO-Local and BC-RNN on Coverage, Pursuit-Evasion, and Box Pushing (Table II), with supporting action- and trajectory-level visualizations (Figs. 3–4).","tokens_in":13861,"tokens_out":1445,"duration_ms":23385,"significance":"If the result holds, the paper makes a clear and useful contribution to multi-robot learning under communication constraints: it identifies mode averaging as a structural failure mode of standard privileged distillation in multi-agent settings, gives a clean formalization of that failure, and shows that conditional diffusion policies are an effective remedy without requiring inter-agent communication or human demonstrations. Strengths include a coherent three-stage pipeline, explicit theory–experiment linkage (Corollary 1 vs. Fig. 3 and BC-RNN collision spikes), controlled ablations (RNN, privileged upper bound, local MARL, deterministic BC), and multi-seed evaluation with task returns rather than training-loss proxies. The work is timely for marine and other bandwidth-limited multi-robot domains and is among the first to combine privileged multi-agent RL with diffusion-based student policies.","major_comments":[{"comment":"The central methodological claim is that multi-modality under partial observability requires modeling a full conditional action distribution rather than a point estimate. The only generative/multi-modal comparison is against deterministic MSE BC-RNN (Table II, §V-C). Without at least one alternative multi-modal student class on the same offline dataset (e.g., a Gaussian mixture / MDN head, CVAE, or simple energy-based policy), it remains unclear whether the gains are due to diffusion’s reverse process specifically or to any capacity to represent multiple modes. Adding one such baseline would substantially strengthen the claim that diffusion is the appropriate tool.","section":"§V-C, Table II; §IV-D"},{"comment":"Stage 2 retains only successful, collision-free oracle episodes and labels them with the deterministic mean of the oracle Gaussian (Algorithm 1 lines 17–19; §IV-B). Multi-modality is then induced by different global configurations that share a local observation, which is valid for the reported BC-RNN vs. CoDiMAD comparison on the same support. However, absolute claims about recovering “coherent coordination modes” and approaching the oracle under decentralized execution rest on that filtered support covering the modes needed at test time. The authors acknowledge offline OOD fragility in §VI; a brief quantitative check (e.g., fraction of evaluation states near the dataset support, or a controlled recovery/perturbation experiment) would make the load-bearing assumption explicit rather than left to the limitation paragraph.","section":"§IV-B, Algorithm 1; §VI"},{"comment":"On Box Pushing, CoDiMAD reaches 72.2% success versus the oracle’s 98.2% and BC-RNN’s 6.5% (Table II). This is the largest residual gap to the privileged upper bound and the setting where force coordination is tightest, yet the paper offers little failure-mode analysis (timing, contact geometry, relative pose ambiguity). Without that analysis it is hard to tell how much of the remaining gap is irreducible partial-observability ambiguity versus distillation or capacity limits, which weakens the interpretation of “approaching oracle-level performance” across tasks.","section":"§V-C, Table II (Box Pushing)"}],"minor_comments":[{"comment":"Proposition 1 and Corollary 1 are clear proof sketches under a well-separated Gaussian-mixture assumption (Δ ≫ σ_max). A short empirical check of mode separation or mixture structure beyond the single Pursuit-Evasion observation in Fig. 3 would better connect the theory to the full dataset.","section":"§IV-D; Fig. 3"},{"comment":"DDIM with K=20 is motivated for real-time control, but no wall-clock latency, control frequency, or onboard timing numbers are reported. Even a brief timing table would support the deployment claim.","section":"§IV-C.3; §I"},{"comment":"Table I and the observation description are helpful; a one-sentence note on sensing radius relative to arena size (200×200) would make the information asymmetry more concrete for readers.","section":"§III-B; Table I"},{"comment":"Related work correctly distinguishes MIMIC-D (human-demonstration multi-agent diffusion) from privileged RL distillation; a slightly sharper sentence on what is new versus CTDS / interactive distillation would help novelty placement.","section":"§II"},{"comment":"Minor presentation: “discard” → “discarded” (§IV-A); arXiv IDs in related work for concurrent preprints are fine but should be updated if camera-ready versions exist; Fig. 1 caption is dense—consider splitting Stage 2/3 labels for readability.","section":"§IV-A; Fig. 1"},{"comment":"Scalability is only evaluated at N=3; the limitation is noted in §VI. A single N=5 or N=6 pilot (even if only Coverage) would reduce the risk that the multi-modality story is specific to small teams.","section":"§V-A; §VI"}],"recommendation":"minor_revision","confidential_remarks":"Solid methods paper for a robotics venue; central claim is internally consistent and the BC-RNN collision/success gaps are convincing evidence for mode averaging. I would not block on the offline-filter limitation alone—it is already disclosed and shared by both distillation methods. The missing multi-modal non-diffusion baseline is the one revision that would most improve the paper’s lasting value. Scope fits multi-robot learning / RAL-style venues well; novelty claim vs. concurrent multi-agent diffusion imitation work is acceptable if phrased carefully."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is simple. Under partial observability the same local map can map to several valid oracle actions; MSE distillation averages them into low-density junk, while a conditional diffusion student can sample a coherent mode. They package that into a three-stage pipeline (MAPPO oracle, successful-episode offline pairs, DDPM/DDIM student) and show consistent gains over local MAPPO and BC-RNN on three continuous cooperative tasks.\n\nWhat is actually new is the multi-agent privileged-distillation framing plus the explicit multi-modality analysis, not any single component. Diffusion policies, MAPPO, and privileged distillation already exist; the contribution is the combination and the clean Corollary 1 diagnosis of why deterministic BC fails. They do that part well. Table II is readable (three seeds × 200 episodes, privileged upper bound, local MARL, BC-RNN, RNN ablation). Fig. 3–4 match the theory: bimodal oracle actions, diffusion covering both modes, BC sitting in the middle, and trajectory diversity only for the diffusion student. The BC collision spikes (especially Coverage) are exactly what mode-averaging predicts. Theory is a short, honest proof sketch under a separated-mixture assumption, not oversold.\n\nSoft spots are real but already flagged by the authors and do not sink the comparisons. Stage 2 keeps only deterministic mean actions from successful, collision-free oracle episodes, so the student never sees recovery behavior outside that support. Same filtered data is used for BC-RNN, so the large gaps (Box Pushing 72 % vs 6.5 %, ~19× fewer collisions on Coverage) still isolate distribution modeling versus mean regression. N=3 only, sim-only, code not yet public, and limited environment/reward detail are the usual methods-paper limits. No hidden communication or reward leakage shows up on a close read.\n\nThis is for people working on communication-free multi-robot RL or multi-modal imitation under partial observability. It deserves a serious referee. I would engage with it, cite the mode-averaging point and the pipeline when relevant, and bring it to reading group if we are talking multi-agent diffusion or privileged distillation.","headline":"Solid compositional methods paper: privileged MAPPO → filtered offline pairs → conditional diffusion students, with a clean mode-averaging diagnosis that the experiments actually support.","tokens_in":14489,"tokens_out":524,"would_cite":true,"duration_ms":7478,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"When robots only see local sensors, the same view can demand several valid joint moves; diffusion distillation samples those modes instead of averaging them into invalid actions.","keywords":["multi-robot systems","privileged policy distillation","partial observability","diffusion policy","multi-agent imitation learning","communication-free coordination","MAPPO","Dec-POMDP"],"falsifier":"On the same three tasks, replace the diffusion student with an otherwise identical deterministic MSE head (or train diffusion on unfiltered oracle samples including failures) and check whether success rates and collisions reverse the gaps reported in Table II, especially the large box-pushing and collision improvements.","tokens_in":14410,"feed_emoji":"🤖","tokens_out":799,"duration_ms":12344,"temperature":0.7,"pith_summary":"This paper argues that communication-free multi-robot teams fail under partial observability because one local sensor reading can match several global situations that require qualitatively different cooperative actions. Privileged distillation first trains a fully informed oracle, then teaches sensor-limited students, but ordinary deterministic cloning collapses those options to their mean and produces hesitant or colliding behavior. CoDiMAD instead models the student as a conditional denoising diffusion policy that approximates the full oracle-action distribution and samples decisive actions from coherent modes. Theory shows why mean-squared regression lands in a low-density gap between modes, while reverse diffusion recovers the mixture. On coverage, pursuit-evasion, and box-pushing tasks with three agents, the method beats both direct local multi-agent RL and deterministic distillation while approaching the non-deployable oracle.","feed_headline":"Diffusion students beat mode-averaging robot distillation","feed_subtitle":"Local views hide multiple valid team moves; sampling modes cuts collisions and lifts success","key_machinery":"CoDiMAD: a three-stage pipeline that trains a privileged MAPPO oracle, builds an offline dataset of local-observation–oracle-action pairs from successful collision-free rollouts, and distills into decentralized conditional DDPM students (accelerated by DDIM) that denoise actions given only local map history.","core_discovery":"Partial observability in multi-agent coordination induces a multi-modal conditional oracle-action distribution; deterministic regression-based distillation therefore converges to the conditional mean, which can lie between valid modes and yield invalid actions, whereas students parameterized as conditional diffusion models approximate that distribution and can sample from individual coordination modes.","pith_inferences":["Any multi-agent setting whose local sensors hide teammate intent (not only marine robots) is a candidate for the same multi-modal failure mode.","If online fine-tuning with diffusion-compatible RL is added, the offline support limitation may shrink enough for physical deployment under distribution shift.","Attention-based denoisers suggested for larger teams would also test whether mode recovery scales when the number of latent global configurations grows combinatorially."],"forward_implications":["Communication-free robot teams can inherit most of a privileged oracle’s coordination skill without runtime message passing.","Deterministic behavioral cloning is systematically unsafe when local observations leave multiple joint strategies open.","Diffusion policies become a practical student class for multi-robot privileged distillation once few-step DDIM sampling is used.","Action-space and trajectory visualizations can diagnose mode averaging versus mode recovery in multi-agent imitation."],"fun_headline_variants":["Diffusion samples multi-modal robot moves that averages collapse","Conditional diffusion recovers modes deterministic robot distillation misses","CoDiMAD diffusion students pick decisive actions mean regression ruins","Mode-averaging fails multi-robot distill; diffusion draws valid coordination","Local-obs diffusion beats mean collapse in communication-free robot teams"],"cache_read_input_tokens":7936,"weakest_assumption_plain":"The student only ever sees filtered mean actions from successful, collision-free oracle episodes, so useful multi-modal coordination is assumed to be fully represented by that offline support and not to require recovery skills outside it.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion samples multi-modal robot moves that averages collapse","Conditional diffusion recovers modes deterministic robot distillation misses","CoDiMAD diffusion students pick decisive actions mean regression ruins","Mode-averaging fails multi-robot distill; diffusion draws valid coordination","Local-obs diffusion beats mean collapse in communication-free robot teams"]},"model":"grok-4.5","effort":"low","cost_usd":0.005512,"raw_usage":{"total_tokens":1482,"prompt_tokens":749,"num_sources_used":0,"completion_tokens":85,"cost_in_usd_ticks":55120000,"prompt_tokens_details":{"text_tokens":749,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":648,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":749,"tokens_out":85,"duration_ms":5373,"temperature":1.0,"reasoning_tokens":648,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T01:56:10.361389+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the same three tasks, replace the diffusion student with an otherwise identical deterministic MSE head (or train diffusion on unfiltered oracle samples including failures) and check whether success rates and collisions reverse the gaps reported in Table II, especially the large box-pushing and collision improvements.","supporting_citations":[],"review_version":1}