{"id":"0168ef48-94f2-4cec-88b1-3b72f291e3a0","arxiv_id":"2607.02977","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"FSD-MP performs diffusion over continuous trajectory functions with Matérn noise and a DST-FNO reverse model, enabling zero-shot multi-resolution robot motion planning.","lead":"Robot motion planners that use diffusion models usually lock to one trajectory resolution and must be retrained for finer or coarser paths. This paper shows a function-space diffusion method that trains once and generates consistent collision-free trajectories at resolutions up to 16 times higher without retraining.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The central claim is empirical and well-supported by Table I (competitive or superior success/valid rates at training resolution, essentially unchanged at 16×) and Fig. 3 (visually identical modes). The function-space construction (mode-wise Matérn SDE + zero-end projection + DST-FNO) is standard and correctly specialized for fixed endpoints; the ablations isolate each design choice. Ordinary robotics-paper limitations (no public code, no error bars, modest environment set) do not undermine the reported result. Consequently the ACCEPT verdict stands; no load-bearing technical concern requires a change.","tokens_in":12781,"tokens_out":404,"duration_ms":3575,"concrete_test":"Re-run the Dense2D and Warehouse ablations of Table II at N=2048 (and, if feasible, N=4096) with the identical trained checkpoint; if FSD-MP valid rates remain within 1–2 points of the N=64 numbers while the FNO and UNet variants continue to collapse, the discretization-invariance claim is further corroborated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest_assumption (Matérn covariance + finite-mode DST-FNO remaining accurate under large grid changes) is the right place to look, but the paper already supplies the necessary support: Appendix B states the trace-class condition α>1/2 for D=1 trajectories, the DST basis is chosen precisely for zero-boundary compatibility (Sec. IV-C), and Table II shows that replacing either the Matérn noise or the DST layers measurably degrades cross-resolution valid rates while the full model stays flat from N=32 to N=1024. No internal inconsistency or hidden assumption that would invalidate the zero-shot claim is present.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes Function-Space Diffusion for Motion Planning (FSD-MP), which treats robot trajectories as continuous functions in a Hilbert space and performs diffusion in the spectral domain rather than on fixed-length waypoint sequences. A mode-wise forward process is driven by Matérn-type Gaussian noise (trace-class for trajectories), and the reverse process is parameterized by a boundary-compatible DST-FNO that, together with a zero-end residual projection and FiLM conditioning, enforces start–goal constraints at any resolution. Classifier guidance on the predicted clean trajectory is used at inference. Experiments on 2D point-robot and 7-DoF Franka benchmarks show that a single model trained at N=64 matches or exceeds fixed-resolution diffusion planners (MPD, MPD-Spline) and GPMP at the training resolution, and generalizes zero-shot to N=1024 (16×) with essentially unchanged success/valid rates and solution modes (Table I, Fig. 3).","tokens_in":13023,"tokens_out":737,"duration_ms":5826,"significance":"Discretization dependence is a genuine practical limitation of existing diffusion planners; a method that trains once and evaluates at arbitrary resolution without retraining is therefore of clear value for robotics. The combination of a well-defined function-space SDE, Matérn noise, and a DST-based neural operator is a clean technical contribution that is supported by ablations (Table II) isolating each design choice. The empirical demonstration of mode-consistent super-resolution up to 16× on both 2D and 7-DoF tasks is strong and falsifiable. The work sits at a natural intersection of function-space generative modeling and motion planning and should be of interest to both communities.","major_comments":[],"minor_comments":[{"comment":"The abstract and introduction claim generalization “up to 16× higher,” yet Fig. 3 also shows N=2000 (≈31×). A single clarifying sentence on the highest resolution that was systematically evaluated would avoid any ambiguity.","section":null},{"comment":"Table I leaves MPD cells blank for N=1024 (as expected for a fixed-resolution baseline). Adding an explicit footnote that MPD cannot be evaluated zero-shot would make the comparison table self-contained.","section":null},{"comment":"Sec. IV-A / Appendix B: the precise numerical values of the Matérn hyperparameters (α, κ, σ) used in the experiments are not listed in the main text; placing them in a short table or appendix would aid reproducibility.","section":null},{"comment":"Eq. (12) and the surrounding paragraph correctly state the trace-class condition α>1/2 for D=1; a brief remark that the same construction extends immediately to higher-dimensional configuration spaces (with the corresponding α>D/2) would strengthen the discussion of future work.","section":null},{"comment":"A few minor typographical inconsistencies appear (e.g., “Matérn-type” vs. “Matérn”, occasional missing spaces around citations). These are purely cosmetic.","section":null}],"recommendation":"accept","confidential_remarks":"The reader’s and skeptic’s assessments align with my own: the central zero-shot claim is empirically well-supported and free of circularity. No load-bearing technical flaw is present. The paper is ready for acceptance; the listed minor points can be handled in production."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real result here is empirical and useful: one model trained only at N=64 keeps essentially the same success/valid rates and solution modes out to N=1024 (and even 2000 in the figures) on both 2-D point robots and 7-DoF Franka, without retraining. Fixed-resolution MPD cannot do that; MPD-Retrain has to be re-fit at every grid; MPD-Spline degrades. That is the contribution worth knowing about.\n\nWhat is new is the combination, not the ingredients. Function-space diffusion (Lim, Kerrigan, Pidstrigach) and FNOs already exist; Diffuser/MPD already do trajectory diffusion. The authors adapt them cleanly for planning: Matérn mode-wise noise so the covariance is actually trace-class, zero-end residual projection so endpoints stay exact, FiLM conditioning, and a DST-FNO whose sine basis respects the boundary. Appendix B states the α>1/2 condition; Table II shows that swapping either Matérn or DST for white noise or plain FFT measurably hurts cross-resolution valid rates while the full model stays flat. The math is standard and consistent; the citation pattern is honest.\n\nSoft spots are ordinary for a robotics methods paper, not load-bearing. Environments are the usual four (Simple/Dense2D, Spheres3D, Warehouse) with the usual RRTConnect+GPMP expert data; no error bars, no public code yet, free parameters (Matérn α/κ/σ, mode count K, guidance weight) are not exhaustively swept. Guidance is the usual classifier-style approximation on the predicted clean trajectory. None of that overturns the zero-shot claim the tables and figures actually show.\n\nThis is for people who already care about generative motion planners or multi-resolution control stacks. It will not reorganize broader ML, but it removes a real engineering bottleneck inside the subfield. I would bring it to reading group, I would cite the multi-resolution numbers if I were working on diffusion planners, and a serious editor should send it to referees rather than desk-reject. Accept after ordinary polishing.","headline":"Solid engineering paper: function-space diffusion + DST-FNO actually delivers zero-shot 16× multi-resolution planning that fixed-grid diffusion planners lack.","tokens_in":13631,"tokens_out":560,"would_cite":true,"duration_ms":4907,"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":"A single diffusion model trained on coarse trajectories can plan robot paths at 16\times finer resolution without retraining.","keywords":["function-space diffusion","motion planning","discretization invariance","Fourier neural operator","DST-FNO","Matérn covariance","robot trajectory generation"],"falsifier":"Train once at resolution 64, then evaluate success and valid rates at resolution 1024 (and intermediate grids) on the same start-goal pairs; if either rate collapses or the generated modes visibly change relative to the training resolution, the zero-shot claim fails.","tokens_in":13671,"feed_emoji":"🤖","tokens_out":785,"duration_ms":6354,"temperature":0.7,"pith_summary":"Robot motion planners that use diffusion models usually treat a path as a fixed list of waypoints, so a model trained at one sampling rate cannot be used at another. This paper treats every trajectory as a continuous function and runs the entire diffusion process in that function space. The result is a planner that can be trained once on a coarse grid and then evaluated, with no extra training, on grids up to sixteen times denser while still producing the same collision-free, multimodal paths. The authors show this works for both planar point robots and a seven-degree-of-freedom arm, matching or beating fixed-resolution diffusion planners at the original resolution and keeping the same success rates at every finer resolution they tried. Readers who care about flexible, high-fidelity planning without having to retrain for every controller or environment resolution will find the claim directly useful.","feed_headline":"One diffusion model plans robot paths at 16\times finer resolution","feed_subtitle":"Train once on coarse waypoints; generate collision-free trajectories on any denser grid without retraining","key_machinery":"Mode-wise spectral diffusion driven by Matérn-type Gaussian noise, reversed by a Discrete Sine Transform-based Fourier Neural Operator (DST-FNO) that is boundary-compatible with fixed start and goal configurations.","core_discovery":"Function-space diffusion for motion planning (FSD-MP) models trajectories as continuous functions and performs diffusion directly in that infinite-dimensional space, so a model trained at one discretization can generate trajectories at arbitrary finer or coarser resolutions without retraining while preserving start-goal constraints and solution modes.","pith_inferences":["The same spectral construction could be applied to other continuous robotic signals such as force profiles or end-effector paths, not only configuration trajectories.","Local refinement near narrow passages could be realized simply by increasing resolution only on selected time intervals, without retraining the operator.","If the Matérn length-scale is made state-dependent, the method may automatically allocate more modes where obstacle density is high."],"forward_implications":["A single trained model can serve controllers that demand different temporal resolutions without any fine-tuning.","Coarse-to-fine or adaptive-resolution sampling becomes a free inference choice rather than a training redesign.","Start-goal constraints remain exactly satisfied at every resolution by construction, removing post-hoc endpoint repair.","The same function-space prior can be reused across 2-D point-robot and 7-DoF manipulator tasks while retaining multimodality."],"fun_headline_variants":["Function-space diffusion plans robot trajectories at any resolution","Diffusion in function space yields resolution-invariant robot paths","One model generates collision-free paths at 16x denser grids","Function-space diffusion enables zero-shot multi-resolution planning","Trajectories as continuous functions allow grid-free motion planning"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The combination of Matérn-type noise and a finite-mode DST neural operator remains a faithful, well-defined model of trajectory functions when the evaluation grid is changed by more than an order of magnitude.","fun_headline_variants_meta":{"raw":{"variants":["Function-space diffusion plans robot trajectories at any resolution","Diffusion in function space yields resolution-invariant robot paths","One model generates collision-free paths at 16x denser grids","Function-space diffusion enables zero-shot multi-resolution planning","Trajectories as continuous functions allow grid-free motion planning"]},"model":"grok-4.5","effort":"low","cost_usd":0.00318,"raw_usage":{"total_tokens":1090,"prompt_tokens":744,"num_sources_used":0,"completion_tokens":66,"cost_in_usd_ticks":31800000,"prompt_tokens_details":{"text_tokens":744,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":280,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":744,"tokens_out":66,"duration_ms":2953,"temperature":1.0,"reasoning_tokens":280,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T05:40:31.248338+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train once at resolution 64, then evaluate success and valid rates at resolution 1024 (and intermediate grids) on the same start-goal pairs; if either rate collapses or the generated modes visibly change relative to the training resolution, the zero-shot claim fails.","supporting_citations":[],"review_version":1}