{"id":"35d3ff4d-e140-4a11-948e-a610a3d00b0e","arxiv_id":"2502.03607","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SMD couples diffusion trajectory generation with an augmented-Lagrangian projection loop, yielding collision-free, kinematically feasible multi-robot plans with 96-100% success on the introduced benchmark.","lead":"SMD adds a projection step that corrects diffusion-generated multi-robot trajectories during sampling, so the final paths respect collision and speed limits even in cluttered spaces. The paper reports near-perfect success rates where previous diffusion planners fail, and introduces a new benchmark with 4,000 multi-robot planning test cases.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stopping criterion in Algorithm 2 as written checks the wrong quantity (multiplier gradient instead of constraint residual), so the claimed feasibility guarantee in Remark 4.3 is not actually established by the stated procedure.","rationale":"I re-derived the projection logic from Section 4.3. The auxiliary variables are defined by the equalities after Eq. (8), but Algorithm 2's only updates are to the trajectory and the penalty weight ρ; the multipliers νa, νo are never updated, and da, do are never updated. So the dual ascent (12a)-(12c) is not implemented as written. The reader's weakest assumption identified exactly this cluster: the relaxation may not enforce the true constraints, and the proof of Proposition 4.2 only covers the convex set. I agree that this is the most load-bearing concern, since the paper's headline claim of feasibility ('perfect success rates and collision ratios') is tied to Remark 4.3. The empirical results are extensive and reported in detail, and the code is released, which is independent support; but the theoretical guarantee and the algorithm pseudocode need correction. The correct verdict is CONDITIONAL, unchanged from the reader, because the empirical claim could survive with a rewritten guarantee, but the paper should not be accepted without fixing the projection procedure and its proof.","tokens_in":19611,"tokens_out":1550,"duration_ms":12374,"concrete_test":"Instrument the released code to log, at every projection call: the max constraint residuals max|Ha(Π)|, max|Ho(Π)|, the actual while-loop condition, and the final residual after termination. Re-run the dense 9-robot and corridor 2-robot experiments. If the final residuals exceed δa or δo, or if the loop exits without performing the minimizations described in Eq. (12a), the Remark 4.3 guarantee does not hold for the code as released, and the central feasibility claim needs to be re-scoped to an empirical statement.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim (96-100% success) is credible, but the theoretical guarantee backing it is not. Remark 4.3 claims the projection output satisfies ||Ha(Π)|| ≤ δa and ||Ho(Π)|| ≤ δo. Yet Algorithm 2's while condition is '∇νa < δa and ∇νo < δo', where ∇νa and ∇νo are set to Ha(Π), Ho(Π). This conflates the dual residual with the primal constraint violation, and the loop as written inverts the intended semantics: if the constraint residuals are initialized below tolerance it exits immediately, and there is no update to the multipliers νa, νo inside the loop. Additionally, the auxiliary variables da, do introduced in Eq. (8) to convert inequalities to equalities are never updated, so it is unclear whether the equalities of Ha, Ho are actually satisfiable. Proposition 4.2 only proves a convex feasibility guarantee (distance to Ωc), and its proof in Appendix B relies on Eq. (15) and Theorem 5.2 from prior work without verifying the theorem's conditions. Thus the paper's formal claim that SMD produces collision-free trajectories within tolerance is not supported by the stated algorithm and proof, even though the reported experiments may be valid.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Simultaneous MRMP Diffusion (SMD), a diffusion-based multi-robot motion planner that alternates denoising steps with a projection operator designed to enforce collision avoidance and kinematic feasibility. The projection is cast as a Lagrangian relaxation over convex constraints (start/goal, velocity) and nonconvex constraints (inter-robot and obstacle avoidance), solved by a dual-ascent procedure (Algorithm 2). The authors claim a feasibility guarantee: the projection output satisfies ||Ha(Π)|| ≤ δa and ||Ho(Π)|| ≤ δo for user-specified tolerances (Remark 4.3). They also introduce a new MRMP benchmark with empty, basic, dense, corridor, shelf, and room maps, and report extensive experiments showing that SMD outperforms standard diffusion (DM), Motion Planning Diffusion (MPD), and Multi-robot Motion Planning Diffusion (MMD), with high success rates and near-zero collision ratios across most settings. Classical baselines (EECBS) are also compared in the appendix.","tokens_in":19748,"tokens_out":4076,"duration_ms":37451,"significance":"If the feasibility guarantee and the experimental results hold, SMD would be an important advance: it would be the first diffusion-based planner to produce collision-free, kinematically feasible multi-robot trajectories in cluttered environments without rejection sampling or post-hoc filtering. The paper ships code, provides a standardized benchmark with 4,000 test instances, and reports consistent, monotonic improvements across all six map types and all robot counts, including a 3.6x success-rate improvement over the prior state of the art in dense 9-robot maps (96% vs. 27%). The sensitivity analysis of the projection scaling factor is a useful addition. However, the theoretical justification is not yet sound: the stated algorithm and proof do not establish the advertised nonconvex feasibility guarantee, and the relaxation with auxiliary variables is not fully specified. The empirical claim is credible on its own, but the paper's central formal contribution—the tolerance-bounded feasibility guarantee—is not supported by the current writeup.","major_comments":[{"comment":"The stopping criterion in Algorithm 2 checks the wrong quantity. The while condition is '∇ν_a < δa and ∇ν_o < δo', but lines 3 and 5 set ∇ν_a and ∇ν_o to H_a(Π̂) and H_o(Π̂), i.e., primal constraint residuals, not the dual gradients. This conflates the dual residual with the primal violation. Moreover, the loop never updates ν_a or ν_o, and line 4 uses ν*_a, ν*_o, which are never defined in the algorithm. As written, the loop terminates immediately if the initial residuals are below tolerance, and it does not implement the dual ascent described in Eq. (12). Consequently, Remark 4.3's claim that the output satisfies ||H_a(Π)|| ≤ δa and ||H_o(Π)|| ≤ δo is not established by the stated procedure.","section":"§4.3, Algorithm 2"},{"comment":"The auxiliary variables d_a and d_o introduced in Eq. (8) to convert inequalities into equalities are never optimized. The Lagrangian in Eq. (9) and the primal update in Eq. (12a) minimize over Π ∈ Ω_c only, with no minimization or update over d_a, d_o. Thus the equality constraints H_a(Π) = 0 and H_o(Π) = 0 as defined in Eq. (8) are not enforced; the residuals computed in Algorithm 2 are not equality residuals of the relaxed system. Without an update rule for d, the relaxation does not correspond to the original nonconvex inequalities, and a trajectory with small H_a(Π) (interpreted as the inequality violation f(Π) − R_a²) may still violate the original constraints, or the equalities may be infeasible for any Π. The paper should either explicitly minimize over d, or justify why the inequality-violation form is sufficient for the claimed guarantee.","section":"§4.3, Eq. (8)-(12)"},{"comment":"The proof of Proposition 4.2 is incomplete and does not support Remark 4.3. Eq. (15) asserts the existence of an iteration index Ī and a bound involving ‖F_t‖ without proof, and the conditions of Theorem 5.2 of Christopher et al. (2024) are not verified in the present setting. More importantly, Proposition 4.2 only bounds dist(Π, Ω_c), the distance to the convex set; it says nothing about the nonconvex constraints H_a(Π) and H_o(Π). Remark 4.3's guarantee on nonconvex residuals therefore does not follow from Proposition 4.2. The paper needs either a complete proof that the dual-ascent procedure drives the nonconvex residuals below δ, or a revised claim that clearly separates the convex guarantee from the empirical/heuristic treatment of nonconvex constraints.","section":"§4.4 and Appendix B, Proposition 4.2"},{"comment":"The claim in Section 6.1 that 'SMD is the only known method that provides feasible solutions for the largest number of robots in complex environments' is an overclaim relative to the evidence. The paper compares SMD only with DM, MPD, MMD, and (in the appendix) EECBS; other learning-based MRMP or optimization-based planners are not included. Similarly, the contribution stating 'the first benchmark for MRMP evaluation' is contradicted by Section C, which evaluates on maps introduced by Shaoul et al. (2025). These statements should be scaled back to 'among the methods compared' and 'a new benchmark,' respectively.","section":"§6.1 and §5, Baseline comparison"}],"minor_comments":[{"comment":"Typo: 'especifically' should be 'especially'.","section":"§2, Related Work"},{"comment":"In the Shelf Maps 9-robot rows, the 'C ↓' and 'A ↓' entries appear swapped: the collision-ratio row contains acceleration values and the acceleration row contains collision-ratio values. This makes Table 4 hard to read and should be corrected.","section":"Table 4, Shelf Maps, 9 robots"},{"comment":"The y-axis of the sensitivity analysis plots is labeled 'Gradient,' but the text refers to 'gradient convergence.' Please clarify whether the plotted quantity is the constraint residual, the multiplier gradient, or the norm of the augmented-Lagrangian gradient, since the distinction matters for interpreting the convergence behavior.","section":"Figures 9 and 10, Appendix C"},{"comment":"Eq. (12c) contains a typo: the update for ν_o is written as 'ν_o^{k+1} = ν_a^k + ρ_o^k H_o(Π^k)', which reuses the subscript 'a' instead of 'o'. The intended update should be ν_o^{k+1} = ν_o^k + ρ_o^k H_o(Π^k).","section":"§4.3, Eq. (12c)"},{"comment":"The nonconvex constraints are written as squared-distance inequalities, but they are not qualified by whether the trajectories are continuous or discretized in time. Please state explicitly whether (6) and (7) are required at every discrete time step h only, or over continuous time, and how the benchmark evaluates collisions between time steps.","section":"§4.2, Eq. (6)-(7)"}],"recommendation":"major_revision","confidential_remarks":"The empirical evaluation is the strongest part of the paper: the results are consistent and the benchmark is a useful community resource. However, the central theoretical claim—that the projection operator guarantees collision-free trajectories within user-defined tolerances—is not established by the current Algorithm 2 and Proposition 4.2. The issues are fixable in principle (correct the stopping criterion, specify the dual updates and the role of auxiliary variables, or weaken the guarantee to an empirical statement), but they affect the paper's central claim. I recommend major revision rather than reject, because the empirical results appear credible and the theoretical gaps are localizable rather than fatal to the overall approach. The authors should also temper the 'only known method' and 'first benchmark' claims to match the actual comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hey, here's the deal with 2502.03607. The empirical story is strong: SMD beats the baselines by a wide margin across all map types, including the hard 9-robot dense maps where MMD collapses to 27%. On sheer results, this is a useful advance. The benchmark of 4,000 instances is also a real contribution; code is up.\n\nBut the theory section does not back up the paper's feasibility claims. Algorithm 2 has a load-bearing bug as written: the while loop continues while \\nabla\\nu_a < \\delta_a, which is backwards (you want to loop until the residual is below tolerance, not while it's below). The multipliers \\nu_a and \\nu_o are never updated inside the loop, and line 4 uses \\nu^*_a,\\nu^*_o that are never defined. The auxiliary variables d in Eq. (8) are introduced but never touched again, so the equality relaxation is incomplete. Proposition 4.2's proof relies on Eq. (15), which is asserted without proof, and on a theorem from your prior NeurIPS paper without checking its conditions. Remark 4.3 then claims the output satisfies ||Ha||≤δa etc., but that's exactly what Algorithm 2 doesn't establish. At best, the proposition covers the convex set \\Omega_c, not the collision constraints.\n\nThe weak theory doesn't kill the empirical result—the trends are monotonic and the success-rate gaps are big—but it means the paper overstates its guarantees. There are also minor issues: no error bars on success rates/path lengths, and calling the benchmark 'the first' is wrong since Shaoul et al. already had practical maps; the paper even evaluates on those later, which undercuts the claim.\n\nWho should read it: people working on learning-based MRMP or constrained diffusion sampling. The method idea—repeated projection with an augmented-Lagrangian dual-ascent loop inside diffusion sampling—is worth taking seriously, but right now the details are too sloppy to reproduce from the text alone.\n\nMy call: this deserves a proper peer review, not a desk reject. The empirical core is valuable and the flaws are fixable in revision. But I'd push for a major revision: rewrite Algorithm 2 correctly, either prove the collision-constraint guarantee under the actual nonconvex constraints or cut the Remark 4.3 claim down to what's proven, and add error bars. If the authors fix the algorithm description and keep the results, it could be a solid publication.","headline":"Real empirical gains, but the feasibility guarantee in the text doesn't hold up because Algorithm 2's stopping condition is inverted and the multipliers never update.","tokens_in":20398,"tokens_out":2801,"would_cite":true,"duration_ms":24201,"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":"SMD projects diffusion sampling into feasible multi-robot trajectories.","keywords":["multi-robot motion planning","diffusion models","constrained optimization","augmented Lagrangian method","collision avoidance","trajectory generation","MRMP benchmark"],"falsifier":"Run SMD on the corridor swap task with tight tolerances ($\\delta_a=\\delta_o=10^{-4}$), then compute the true collision inequalities $(\\pi_i^h-\\pi_j^h)^2 \\ge (R_a)^2$ and $(\\pi_i^h-o_j)^2 \\ge (R_o)^2$ directly; if any pair of robots or a robot and an obstacle violates these while the residual check reports success, the stopping criterion is not a certificate of collision-freeness.","tokens_in":19266,"feed_emoji":"🤖","tokens_out":5453,"duration_ms":46801,"temperature":0.7,"pith_summary":"This paper tries to prove that diffusion models can plan for multiple robots simultaneously, not just generate candidate trajectories, if the sampling loop is interleaved with a constrained-optimization projection. It introduces SMD, which alternates score-based denoising with an augmented-Lagrangian project-and-dual-update step, and claims this yields collision-free, kinematically feasible plans where gradient-guided and rejection-sampling diffusion planners fail. The reported experiments back this with near-perfect success rates across a new 4,000-instance benchmark, including 96% success in the hardest dense 9-robot maps. If the claim holds, learning-based planners become a practical option for high-density, unstructured multi-robot environments.","feed_headline":"Projected diffusion keeps multi-robot plans collision-free in dense maps","feed_subtitle":"Constrained projection during sampling yields feasible trajectories where other learning-based planners fail, including 96% success with 9…","key_machinery":"The central object is the projected diffusion sampler (Algorithm 1) combined with an augmented-Lagrangian dual-ascent projection (Algorithm 2). The projection maps a noisy trajectory to the nearest point in the convex set $\\Omega_c$ (start, goal, and velocity limits) while iteratively raising Lagrangian multipliers on the nonconvex inter-robot and obstacle-avoidance residuals $H_a$ and $H_o$, with tolerances $\\delta_a$ and $\\delta_o$ as stopping criteria. The relaxation replaces inequality constraints by equalities using nonnegative auxiliary variables $d$, which the paper never updates; the theoretical guarantee (Proposition 4.2) covers distance to $\\Omega_c$, while Remark 4.3 extends the guarantee to the collision residuals only through the user-defined stopping criterion.","core_discovery":"The paper introduces SMD, a diffusion-based multi-robot motion planner that alternates score-based denoising with a projection operator; each projection solves a Lagrangian relaxation of the MRMP constraints over a convex set of start, goal, and velocity-feasible trajectories, then updates dual variables on the nonconvex collision residuals. The central claim is that this produces trajectories satisfying $\\|H_a(\\Pi)\\| \\le \\delta_a$ and $\\|H_o(\\Pi)\\| \\le \\delta_o$ for user-defined tolerances, and empirically yields collision-free, kinematically feasible plans at success rates no other tested learning-based method reaches, including 100% success on corridor swaps, 96% success with 9 robots in dense maps, and a 3.6x success-rate improvement over the prior best in the hardest setting.","pith_inferences":["If the residual bounds are genuine certificates, the same projection wrapper could be dropped onto any generative trajectory model, not only score-based diffusion, turning them into constrained planners without retraining.","The auxiliary variables $d$ in Eq. (8) are never updated in Algorithm 2; whether the stopping criterion can be fooled by stale $d$ is an open stress test the paper does not run.","The runtime growth (from about 30 seconds for 3 robots to about 550 seconds for 9 robots in dense maps) suggests the dual-ascent projection, not the diffusion sampling, is the bottleneck; warm-started multipliers or learned initialization could be a natural next step.","The 4,000-instance benchmark, with corridor, shelf, and room maps, looks usable as a shared evaluation suite for future MRMP planners, including classical anytime planners compared on the same metrics."],"forward_implications":["In dense maps with 20 obstacles and 9 robots, SMD reports 96% success, a 3.6x improvement over the best prior method, with a near-zero collision ratio.","In corridor maps, SMD is the only tested method with any successful plans, achieving 100% success on the two-robot swap.","The projection guarantees output within user-set tolerances $\\delta_a$ and $\\delta_o$ for nonconvex collision residuals, and arbitrarily small distance to the convex constraint set $\\Omega_c$.","The introduced benchmark gives 4,000 standardized instances across empty, basic, dense, corridor, shelf, and room maps, with success rate, path length, acceleration, and collision ratio metrics.","Because no rejection sampling or post-hoc repair is needed, SMD's per-sample pipeline is a single forward pass with interleaved projections."],"supporting_citations":[{"why":"Supplies the augmented Lagrangian and dual ascent machinery used in Algorithm 2.","marker":"Boyd et al., 2011"},{"why":"Supplies the repeated-projection diffusion framework and the Theorem 5.2 bound used in the proof of Proposition 4.2.","marker":"Christopher et al., 2024"},{"why":"The motion-planning diffusion baseline that SMD extends to multi-robot settings and outperforms.","marker":"Carvalho et al., 2023"},{"why":"The multi-robot diffusion baseline and the practical maps used for comparison in Table 6.","marker":"Shaoul et al., 2025"},{"why":"Provides the convergence guarantee for Langevin-style sampling with projections that underpins the constrained diffusion argument.","marker":"Xu et al., 2018"},{"why":"Justifies converting inequality constraints to equalities with nonnegative auxiliary variables in Eq. (8).","marker":"Kotary & Fioretto, 2024"}],"fun_headline_variants":["Projected diffusion solves multi-robot motion planning","Diffusion plus projection clears robot collisions","SMD: simultaneous multi-robot planning with diffusion","96% success in dense maps via projected diffusion","Constrained diffusion sampling for multi-robot paths"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the augmented-Lagrangian dual-ascent loop, which only minimizes over the convex set $\\Omega_c$, actually drives the original nonconvex collision inequalities to zero, so that a small residual means a genuinely collision-free trajectory.","fun_headline_variants_meta":{"raw":{"variants":["Projected diffusion solves multi-robot motion planning","Diffusion plus projection clears robot collisions","SMD: simultaneous multi-robot planning with diffusion","96% success in dense maps via projected diffusion","Constrained diffusion sampling for multi-robot paths"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1168,"prompt_tokens":865,"completion_tokens":303,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":232}},"tokens_in":481,"tokens_out":303,"duration_ms":3950,"temperature":1.0,"reasoning_tokens":232,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T04:21:48.950481+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SMD on the corridor swap task with tight tolerances ($\\delta_a=\\delta_o=10^{-4}$), then compute the true collision inequalities $(\\pi_i^h-\\pi_j^h)^2 \\ge (R_a)^2$ and $(\\pi_i^h-o_j)^2 \\ge (R_o)^2$ directly; if any pair of robots or a robot and an obstacle violates these while the residual check reports success, the stopping criterion is not a certificate of collision-freeness.","supporting_citations":[{"cited_title":"T., Baierl, M., Koert, D., and Peters, J","cited_arxiv_id":null,"evidence_quote":"The motion-planning diffusion baseline that SMD extends to multi-robot settings and outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The multi-robot diffusion baseline and the practical maps used for comparison in Table 6."}],"review_version":1}