{"id":"fb2014be-4bcb-4ae5-93a8-f8ab624a0e63","arxiv_id":"2504.17534","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper proposes combining multidimensional scaling with road-network graph embeddings as feature spaces for generalizable autonomous driving motion planning, but provides only a literature review and toy visualizations, no validated results.","lead":"This paper suggests that road networks, drawn as graphs, can be embedded into map-like feature spaces using a classic distance-preserving technique called multidimensional scaling, to help self-driving cars generalize across different road layouts. It reviews the idea and shows small toy examples, but does not test it on real driving tasks.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim requires MDS to preserve directed, asymmetric travel times of road networks, but the paper's stress formulation (Eq. 1) and toy examples assume symmetric distances, leaving the key structure unhandled.","rationale":"The reader's rejection is based on the absence of empirical validation and the untested distortion assumption. Our concern is more specific: the proposed MDS formulation is inherently symmetric, while road networks are directed with asymmetric travel times. This is not just a missing experiment; it is a potential internal mismatch between the problem setting and the method. The paper itself notes in Section III that projecting time-space onto two dimensions cannot be as accurate as physical space, but it does not address the directed/asymmetric case at all. The toy examples (Figures 9 and 10) are undirected, so they cannot reveal this gap. If asymmetric distances cannot be embedded with acceptable fidelity, the central claim that MDS-based feature spaces improve generalization for motion planning is unsupported even in principle. Our concrete test would isolate this issue: it would show whether the directional information is preserved or lost, and whether that loss matters for a downstream task. Since this concern reinforces the reader's rejection rather than changing it, the verdict remains REJECT, and we mark agreement as partial because the reader identified the broader distortion issue but not the specific asymmetry gap.","tokens_in":17574,"tokens_out":2930,"duration_ms":30154,"concrete_test":"Implement the proposed MDS pipeline on a small directed road network, e.g., a 3x3 grid with one-way streets, using asymmetric shortest-path travel times as input. Compute the stress separately for ordered pairs (i,j) and (j,i) after embedding; if the embedding is forced to be symmetric, measure the directional error (e.g., average |d_ij - ||x_i - x_j||| for ordered pairs vs. the symmetric stress). Then compare a downstream task, such as predicting pairwise travel times or interaction potential, using the MDS coordinates versus the original graph distances. If the embedding loses directional information to a degree that degrades downstream predictions, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the central claim to hold, MDS-based embeddings must preserve the travel-time structure that determines vehicle interaction potential. Road networks are directed: one-way streets, turn restrictions, and traffic flow create asymmetric distances d_ij != d_ji. The paper's stress function (Eq. 1) sums over all i!=j with a single distance d_ij, and the optimization examples (Figures 9 and 10) are undirected trees and pentagons. Section III.B mentions asymmetric proximities only in passing and never integrates them into the optimization or the proposed pipeline. In any Euclidean or kappa-stereographic space, distances are symmetric, so exact preservation of asymmetric travel times is impossible. The paper does not argue that approximate preservation of directional information is sufficient, nor does it propose an asymmetric MDS variant for road networks. Without this, the resulting feature space could misrepresent the interaction potential between vehicles (e.g., treating a one-way street as bidirectional), which would undermine the claimed generalization benefit for motion planning. The absence of any road-network experiment means this gap is entirely unaddressed, so the paper's central claim rests on an untested, and arguably false, assumption that symmetric MDS can serve as a feature space for directed road networks.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes using multidimensional scaling (MDS) to embed road-network graphs into low-dimensional Euclidean or kappa-stereographic spaces in which pairwise distances approximate travel times rather than Euclidean distances, with the goal of providing better feature spaces for learning-based motion planning in autonomous driving. It surveys graph-extraction methods for road networks, reviews classical and gradient-based MDS variants, discusses node-embedding approaches, and illustrates the idea with the optimization of a 16-node tree and a pentagon in a kappa-stereographic space with learnable curvature.","tokens_in":17805,"tokens_out":2859,"duration_ms":28068,"significance":"The paper identifies a real limitation of current practice: using Euclidean coordinates as a proxy for vehicle interaction potential is misleading in road networks where travel time is the relevant metric. If the proposed MDS-based embeddings could be computed with low distortion for real road networks, they would provide a principled, topology-aware input representation for downstream prediction and motion planning models and could improve generalization across road structures. The paper also provides a useful review of MDS methods and graph representations relevant to autonomous driving. However, the central claim is not validated: there are no experiments on road-network graphs, no quantification of embedding distortion, no comparison with baseline representations, and no downstream planning or prediction evaluation. The two toy examples are too small and too far from the road-network setting to support the proposed approach, and the directed, asymmetric nature of road travel times is acknowledged but never incorporated into the optimization or pipeline. As it stands, the paper is best viewed as a position or survey document rather than a demonstration of a working method.","major_comments":[{"comment":"The only experimental evidence is the optimization of a 16-node tree and a pentagon, neither of which is a road network graph. The stress function in Eq. (1) uses symmetric distances d_ij, and both toy examples are undirected. Road networks, however, are directed and have asymmetric travel times due to one-way streets, turn restrictions, and traffic flow. The paper mentions asymmetric proximities in passing in Section III.B but never integrates them into the optimization or the proposed pipeline. Consequently, the central claim that MDS embeddings can serve as feature spaces for road networks is not supported by the presented evidence.","section":"Section IV, Figures 9 and 10; Eq. (1)"},{"comment":"The paper acknowledges that 'it is not possible to project time-space onto a two-dimensional plane as accurately as with physical space' (text immediately preceding the discussion of isochrones). This admitted distortion is never quantified, nor is it argued to be acceptable for the downstream task. Without an error analysis on actual road networks, such as stress values, comparison with Euclidean-distance baselines, or sensitivity to traffic conditions, the reader cannot judge whether the proposed embeddings preserve the interaction-relevant structure needed for motion planning.","section":"Section I, time-space mapping paragraph"},{"comment":"The described pipeline—graph extraction, node embedding, MDS optimization, and use as features for downstream prediction or motion planning—is not implemented or evaluated. The paper asserts that these embeddings can be used as inputs for learning-based motion planning, but no experiment demonstrates this. As a result, the title and abstract overstate what is actually established, and the claimed generalization benefit for autonomous driving remains untested.","section":"Section III.D"},{"comment":"The convergence and stress comparisons shown in Figures 6-8 are reproduced from Zheng et al. on the SuiteSparse matrix collection, not on road-network graphs. These figures may illustrate general properties of SGD versus majorization, but they do not address the scale, sparsity, or directed nature of road networks, so they do not provide direct evidence for the paper's application setting.","section":"Section III.C, Figures 6-8"}],"minor_comments":[{"comment":"The caption says the pentagon 'assumes a spherical form because of the hyperbolical space it is being optimized on,' yet the reported curvature κ takes both negative and positive values (e.g., κ=1.367 in step (h)), so the embedding space is not purely hyperbolic; the caption misdescribes the geometry.","section":"Figure 10 caption"},{"comment":"The text and figure captions refer to 'SDG' when the intended method is stochastic gradient descent (SGD); please correct the abbreviation for consistency.","section":"Figures 6, 7, 8 and Section III.C"},{"comment":"The phrase 'the composableCommonRoad' should read 'the composable CommonRoad' benchmarks, and the manuscript contains several spacing and typographical errors (e.g., 'V ectornet', 'V ehicles', 'disatnces', 'defined') that should be corrected.","section":"Section II.A"},{"comment":"The citation for Google Trends data points to a ResearchGate figure rather than the original source; a direct reference to Google Trends or an archived dataset would be more appropriate.","section":"Reference [4]"},{"comment":"The sentence 'Traffic protocols would have to defined so that the general traffic rules can be preserved' is missing 'be' before 'defined' and should be corrected.","section":"Section IV"}],"recommendation":"reject","confidential_remarks":"The manuscript reads as a survey/position paper rather than a research contribution with a validated method. The novelty claim—that this is the first combination of MDS with learning for autonomous driving—is not substantiated by experiments, and the technical gap between symmetric MDS and directed road networks is fundamental. If the journal accepts position papers as a category, the submission would benefit from reframing; as it stands, the title and abstract promise more than the content delivers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a quick look if you are curious about MDS as a feature-space idea for driving, but this is not a research result. It reads like a well-written position/survey: it lays out why Euclidean distance misrepresents interaction potential on road networks, walks through graph extraction and MDS methods, and ends with two toy optimizations (a tree and a pentagon) in a stereographic embedding space with learnable curvature. Those figures show the mechanics work, but they do not show that MDS can embed a road network usefully for motion planning.\n\nCredit where due: the problem framing is clear, the survey is reasonably current and accurate, and the paper correctly notes that projecting time-space into 2D cannot be as accurate as physical space. The explicit choice of non-Euclidean target spaces with learnable curvature is a reasonable direction, and the toy plots at least demonstrate the optimization pipeline. As an introduction to MDS for graph drawing, it is readable.\n\nThe soft spots are substantial. There is no road-network graph anywhere in the experiments. No downstream planning or prediction evaluation, no comparison against raw Cartesian coordinates or other graph embedding baselines, no quantitative distortion measure on realistic road graphs. The central claim is untested.\n\nThe more specific concern is the directedness of road networks. Eq. (1) and the toy examples use symmetric distances d_ij = d_ji. Road networks have one-way streets, turn restrictions, traffic-dependent asymmetries. In any Euclidean or stereographic space, distances are symmetric, so exact preservation is impossible. The paper mentions asymmetric proximities in Section III.B but never integrates them into the optimization or the pipeline. It never argues that approximate symmetric preservation retains the directional information that matters for vehicle interaction potential. That is a load-bearing gap, and the absence of any realistic experiment means it is completely unaddressed.\n\nThe priority claim in Section I-B (“to the best of our knowledge, no previous attempt…”) is not backed by a systematic search, and is probably too strong given the existence of hyperbolic graph embeddings for road maps and the cited work on time-distance maps. Minor point: no code or data for the figures is provided, so the toy results are not reproducible as shipped.\n\nIf I were a serious editor, I would desk-reject this as a research result; there is no result to evaluate. As a workshop brief or a position write-up, it could generate discussion, and I would not discourage a student from reading it for an MDS refresher and a few good pointers.","headline":"A clearly written position piece that proposes MDS embeddings for road networks but never tests the idea on a road network, and the directedness gap alone keeps it from being a research result.","tokens_in":18322,"tokens_out":2866,"would_cite":false,"duration_ms":27087,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MDS embeds road networks into travel-time feature spaces","keywords":["road networks","multidimensional scaling","time-distance maps","graph embeddings","autonomous driving","motion planning","generalization","non-Euclidean embeddings"],"falsifier":"Embed a real urban road network with travel-time edge weights and measure the normalized MDS stress; if the stress stays high even in three dimensions, or if a motion planner trained on the resulting coordinates does not beat one trained on Cartesian coordinates, the proposal fails.","tokens_in":17361,"feed_emoji":"🚗","tokens_out":12104,"duration_ms":98551,"temperature":0.7,"pith_summary":"The paper seeks to improve generalization in learning-based autonomous driving by replacing naive Euclidean scene coordinates with embeddings of road-network graphs in which distances mean travel time. It argues that multidimensional scaling (MDS) applied to graphs whose edge weights are travel times produces time-distance maps: coordinate sets whose inter-node distances reflect how long travel actually takes, not physical proximity. The author reviews how road scenes can be turned into graphs and which families of MDS (classical, stress-majorization, gradient-based) best minimize the stress between given and embedded distances. Two illustrative experiments, a 16-node tree and a pentagon, show that such embeddings can be optimized in non-Euclidean spaces with learnable curvature, driving stress close to zero. If the approach holds on real road networks, planners trained on these embeddings would ignore geometry that does not affect travel time and generalize across unseen road topologies.","feed_headline":"MDS embeds road networks into travel-time feature spaces","feed_subtitle":"Travel-time-aware coordinates could help motion planners generalize across unseen road topologies.","key_machinery":"The load-bearing construction is the weighted stress function $stress(X)=\\sum_{i\\ne j} w_{ij}\\bigl(d_{ij}-\\|X_i-X_j\\|\\bigr)^2$, where $d_{ij}$ is the shortest-path travel time between graph nodes and $w_{ij}=d_{ij}^{-\\alpha}$ with the common choice $\\alpha=2$. Minimizing this stress yields a time-distance map, an embedding in which pairwise distances approximate travel times rather than physical road lengths. For the paper's demonstrations the optimization runs in a $\\kappa$-stereographic space whose curvature is learned alongside the node coordinates, so the geometry of the embedding adapts to the graph. This machinery converts a road-network graph into a small set of coordinates that a neural planner can use directly.","core_discovery":"The paper's central claim is that graph representations of road networks, with pairwise travel times as distances, can be embedded into a low-dimensional feature space by multidimensional scaling, giving learning-based motion planning a representation insensitive to irrelevant geometric variation. The resulting time-distance map deforms physical space so that node distances track travel time, and a weighted stress function is minimized over node coordinates. The author further claims that performing the optimization in a non-Euclidean space with a learnable curvature parameter can drive the distortion nearly to zero on simple graphs, and that the resulting node coordinates serve as compact, topology-aware inputs for downstream learning, mapping nodes with equal travel distances to the same location no matter their physical origin. The presentation is explicitly exploratory: the evidence is qualitative and limited to a tree and a pentagon, and the author concedes that projecting time-space onto two dimensions cannot be as accurate as projecting physical space.","pith_inferences":["If the embedding is trained jointly with the downstream planner, the same stress function could act as a differentiable regularizer that keeps the latent space faithful to travel-time distances, a step the paper does not explicitly propose.","A stronger test than the paper's examples would compare MDS-based embeddings with graph neural network encoders on the same motion-planning benchmark; the MDS representation should win when preserving global travel-time geometry matters more than local structure.","On real road networks, a learnable curvature may encode systematic travel-time anisotropies such as one-way streets, elevation, and congestion patterns, but could also overfit the training topology if the curvature is fit per city.","The invariance argument suggests a transfer-learning path: a planner trained on one city's time-distance map might transfer to another city whose travel-time geometry is similar even if the physical geometry is different."],"forward_implications":["Motion planners trained on time-distance map coordinates should be insensitive to physical layout details that do not change travel time, such as exactly where an oncoming lane begins.","Replacing raw scene coordinates with MDS node coordinates reduces the dimensionality of the input, which can ease learning and speed up inference.","Because stress minimization emphasizes local distances, the embeddings are well suited to capturing local road structure such as merges, intersections, and one-way constraints.","Gradient-based MDS converges to low stress in fewer iterations than majorization, making it a practical candidate for embedding road graphs online during deployment.","Allowing the curvature of the target space to be learned adds flexibility: spherical or hyperbolic geometries can represent travel-time patterns that flat Euclidean space distorts."],"supporting_citations":[{"why":"Defines the graph representation of road networks (vertices, directed edges, entry and exit sets) that MDS operates on.","marker":"[37]"},{"why":"Introduces time-distance maps, the feature-space representation the paper proposes.","marker":"[42]"},{"why":"Establishes that projecting time-space onto two dimensions introduces unavoidable distortion, the limitation the paper concedes.","marker":"[43]"},{"why":"Supports the choice of shortest-path distances and the inverse-distance weighting in the stress function.","marker":"[39]"},{"why":"Sets the weighting exponent to 2 in the stress function.","marker":"[40]"},{"why":"Presents stress majorization, the main deterministic MDS optimization technique the paper reviews and compares.","marker":"[99]"},{"why":"Presents stochastic gradient descent for graph drawing, which the paper identifies as reaching low stress fastest and motivates the gradient-based MDS choice.","marker":"[105]"},{"why":"Provides the encoder-decoder framework for graph node embeddings that MDS embeddings plug into.","marker":"[111]"},{"why":"Supplies the manifold-aware optimization library used for the illustrative tree and pentagon embeddings with learnable curvature.","marker":"[130]"}],"fun_headline_variants":["MDS maps roads by travel time, not distance","Isometric road embeddings from MDS travel-time","Travel-time MDS: new space for road graphs","Road network embedding via MDS on time","MDS creates travel-time feature spaces for drives"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire approach depends on travel-time distances surviving the move into a compact map with little error, a step the paper demonstrates only on a 16-node tree and a pentagon.","fun_headline_variants_meta":{"raw":{"variants":["MDS maps roads by travel time, not distance","Isometric road embeddings from MDS travel-time","Travel-time MDS: new space for road graphs","Road network embedding via MDS on time","MDS creates travel-time feature spaces for drives"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000208,"raw_usage":{"total_tokens":1360,"prompt_tokens":855,"completion_tokens":505,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":433}},"tokens_in":471,"tokens_out":505,"duration_ms":5570,"temperature":1.0,"reasoning_tokens":433,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:37:43.543042+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Embed a real urban road network with travel-time edge weights and measure the normalized MDS stress; if the stress stays high even in three dimensions, or if a motion planner trained on the resulting coordinates does not beat one trained on Cartesian coordinates, the proposal fails.","supporting_citations":[{"cited_title":"Gansner, Y ehuda Koren, and Stephen North","cited_arxiv_id":null,"evidence_quote":"Presents stress majorization, the main deterministic MDS optimization technique the paper reviews and compares."},{"cited_title":"Graph Drawing by Stochastic Gradient Descent","cited_arxiv_id":"1710.04626","evidence_quote":"Presents stochastic gradient descent for graph drawing, which the paper identifies as reaching low stress fastest and motivates the gradient-based MDS choice."},{"cited_title":"K-stereographic projection model., 2018","cited_arxiv_id":null,"evidence_quote":"Supplies the manifold-aware optimization library used for the illustrative tree and pentagon embeddings with learnable curvature."}],"review_version":1}