{"id":"af2dad4b-279b-48a4-afdf-32aec13870b0","arxiv_id":"2608.11580","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A new coarse-to-fine system generates complete, lane-level road maps from scratch for driving simulation, reporting 99.8% reachability and 0.24 m endpoint alignment.","lead":"RoadWeaver is a three-stage pipeline that generates large lane-level driving maps from scratch, first creating a road layout, expanding it into a connected network, and then building lane geometry. It reports near-complete connectivity and very small endpoint misalignment, and the authors say the maps export directly into OSM and OpenDRIVE for simulators.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline connectivity and endpoint-alignment numbers are measured after RoadWeaver's own repair and filtering steps, so the metrics are by-construction and do not by themselves establish generation quality or simulation-readiness.","rationale":"The reader's weakest assumption correctly identifies the self-referential endpoint alignment metric and the absence of closed-loop validation. My stress-test sharpens this: the same by-construction problem extends to LCC, reachability, and dead-end ratio, because Section III-C explicitly removes disconnected fragments and reconnects dangling endpoints before the metrics in Table II are computed. This is the most load-bearing weakness because it undermines the quantitative basis for the central claim, not merely the downstream usability claim. That said, an audit could rescue the paper: if the metrics are similar before and after the repair passes, then the generation process itself is doing the work, and the numbers would be meaningful. The import and route-planning results provide some evidence of practical deployability, so the paper is not clearly wrong. I therefore keep the reader's CONDITIONAL verdict: the paper should be accepted only if the repair-stage audit is run and closed-loop driving evaluation is added, with code and data released for independent verification. My concern does not move the verdict because the reader's conditional already anticipates these requirements; it does, however, make them more urgent.","tokens_in":9407,"tokens_out":6710,"duration_ms":68432,"concrete_test":"Recompute the topological and geometric metrics at intermediate stages of Algorithm 1 on the same 100 generated maps: (i) immediately after ADDSUCCESSORLINKS (line 11), before REPAIRLANEGEOMETRY and REPAIRTOPOLOGY; (ii) after REPAIRLANEGEOMETRY only; (iii) after REPAIRTOPOLOGY (final). If the pre-repair reachability is already near 99.8% and the pre-repair endpoint alignment error is already near 0.24 m, the repair-artifact concern is refuted; if the final metrics improve sharply (e.g., endpoint error drops from >1 m to 0.24 m, or reachability rises from <90% to 99.8%), the headline numbers are produced by the repair pass and cannot be cited as evidence of intrinsic generation quality or of superiority over baselines without the same post-processing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Table II: LCC 99.9%, reachability 99.8%, dead-end 10.7%, and endpoint alignment 0.24 m. These numbers are computed on the output of Algorithm 1, which includes REFINEROADGRAPH, REPAIRLANEGEOMETRY, and REPAIRTOPOLOGY (lines 1, 12, 14), and on a road graph that has already undergone A*-based reconnection of dangling endpoints and largest-connected-component filtering in Section III-C. High LCC and reachability and a low dead-end ratio are therefore enforced by the pipeline itself, and the 0.24 m endpoint alignment error is the residual after the repair steps snap or reconnect mismatched lane endpoints. The endpoint metric is defined on connections that RoadWeaver itself 'expects' after ADDSUCCESSORLINKS and then repairs, so it is not an independent measure of lane-level generation quality. The comparison with baselines (4.32–8.17 m) is apples-to-oranges because the baselines lack equivalent repair passes. Section IV-E's only downstream validation is 100% import success and 98.7% route-planning success in Tactics2D; no closed-loop driving, traffic simulation, or behavioral evaluation is run, so the central 'simulation-ready' claim is supported only by format compatibility and shortest-path reachability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RoadWeaver, a coarse-to-fine framework for generating large-scale lane-level HD maps from scratch for autonomous driving simulation. The pipeline consists of three stages: global road skeleton generation using a VQ-VAE and a conditional masked Transformer; skeleton-guided road graph expansion via a structure tensor field and procedural growth with A*-based reconnection; and lane-level HD map construction through a sequence of graph refinement, lane assignment, junction connection, and geometric/topological repair operations (Algorithm 1). The authors report LCC of 99.9%, reachability of 99.8%, dead-end ratio of 10.7%, cycle ratio of 85.2%, endpoint alignment error of 0.24 m, generation time of 1.39--3.50 s, 100% import success in Tactics2D, and 98.7% route-planning success over 1,000 routing tasks. The central claim is that RoadWeaver is the first method to combine from-scratch generation, global topology, lane-level detail, large-scale output, controllability, and direct simulator deployability.","tokens_in":9689,"tokens_out":3898,"duration_ms":40941,"significance":"If the central claim is fully validated, RoadWeaver would be a useful contribution: it offers a single trainable pipeline that goes from a learned global layout to simulator-importable lane-level maps with explicit density control, potentially enabling scalable closed-loop evaluation of autonomous driving systems. The integration with Tactics2D and the explicit report of runtime, connectivity, and downstream routing success are strengths. However, the headline metrics are computed on outputs that have already been filtered and repaired by the method itself, and the only downstream validation is route planning; no closed-loop driving or traffic simulation is performed. Thus the significance of the results, especially the 'simulation-ready' claim, is currently supported only conditionally. The paper also ships no code at review time and omits many training and procedural hyperparameters, which limits reproducibility and makes the robustness of the reported numbers difficult to assess.","major_comments":[{"comment":"The headline topological and geometric metrics (LCC 99.9%, reachability 99.8%, dead-end 10.7%, endpoint alignment 0.24 m) are measured on road graphs that have already undergone A*-based reconnection of dangling endpoints, largest-connected-component filtering (Section III-C), and lane-level repair passes REPAIRLANEGEOMETRY and REPAIRTOPOLOGY in Algorithm 1. Consequently these metrics partly measure the method's own repair objective rather than the quality of the raw generated structure. To support the claim that RoadWeaver 'achieves' these numbers as a generation method, please report the same metrics before each repair/filtering stage, or provide an ablation that isolates the contribution of each repair step. Without this, the numbers in Table II are not an independent measure of generation quality.","section":"Section IV-C / Table II / Algorithm 1"},{"comment":"The endpoint alignment error is defined as the average Euclidean distance between lane endpoints that are 'expected to be connected,' but the expected connections are generated by RoadWeaver's own ADDSUCCESSORLINKS step and then actively repaired by REPAIRLANEGEOMETRY and REPAIRTOPOLOGY. This makes the metric self-referential: the 0.24 m result is the residual after snapping or reconnecting endpoints that the method itself decided should be connected. Please define expected connections from an independent source (e.g., OSM ground truth, human annotations, or a separate verification process) or clearly state that this is a measure of internal consistency after repair, not of lane-level geometric fidelity.","section":"Section IV-C / endpoint alignment error definition"},{"comment":"The claim that generated maps are 'directly deployable' and 'simulation-ready' rests on 100% import success (a format-compatibility check) and 98.7% route-planning success in Tactics2D. Import success shows that the file format is accepted, and routing success shows that a shortest path exists in the generated lane graph, but neither demonstrates that the maps support closed-loop driving, traffic simulation, or behavioral evaluation. Task completion, collision avoidance, and traffic-rule compliance are not tested. Please add at least one closed-loop driving or traffic-simulation experiment (e.g., a simple rule-based agent following a route in a generated map, with metrics such as route completion rate and collision rate), or explicitly restrict the claim to import and routing compatibility rather than full simulation readiness.","section":"Section IV-E"},{"comment":"The comparison with MetaDrive, RoadGen, and HDMapGen in Table II is apples-to-oranges because the baselines are not given the equivalent repair, reconnection, and largest-connected-component filtering passes that RoadWeaver applies. The large differences in dead-end ratio and endpoint alignment error may therefore reflect the presence or absence of post-processing rather than inherent generation quality. Either apply the same repair pipeline to the baselines (if applicable) or report RoadWeaver's metrics without its repair stages, so that the comparison is fair.","section":"Section IV-C / Table II / baseline comparison"}],"minor_comments":[{"comment":"Please report the VQ-VAE latent grid size, codebook size, transformer depth and width, training iterations, and conditioning encoding details; these are needed to reproduce the skeleton-generation stage.","section":"Section III-B"},{"comment":"The t-SNE visualization of driving behavior would benefit from a description of how trajectory features were extracted, which datasets were included, and how the color coding corresponds to scenario types; currently the figure is difficult to interpret.","section":"Figure 1"},{"comment":"The density controllability demonstration shows measured densities close to targets, but there is no quantitative error metric or variance across seeds; consider reporting mean absolute error and standard deviation for density control.","section":"Section IV-B / Figure 4"},{"comment":"The statement that larger HDMapGen graphs 'provide limited quality improvement' is not quantified; either add a quantitative support or remove the claim.","section":"Section IV-D"},{"comment":"Please clarify how the 10 routing tasks per map were sampled (fixed seeds? random? endpoints chosen uniformly at random?) and report the variance of routing success across the 100 maps.","section":"Section IV-E"},{"comment":"The metrics use different definitions of 'node' (intersections, endpoints, turning transition points); please state explicitly how each metric's node count is computed so that cross-method comparison is meaningful.","section":"Table II"},{"comment":"The free parameters sigma, w_tensor, w_inertia, theta_max, delta_s, the snapping radius, and the neighborhood radius are never given concrete values; listing them in a table would greatly aid reproducibility.","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is plausible and the integration with Tactics2D is a positive step. My main concern is that the strongest quantitative claims are measured after the method repairs its own output, making them partially by-construction, and the downstream validation stops at route planning. I recommend major revision: the authors should report pre-repair metrics, make the baseline comparison fair, and add a closed-loop validation. I would also urge the editor to weigh the absence of code and unreported hyperparameters when assessing reproducibility; the 'code will be released upon acceptance' statement does not help the review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RoadWeaver is a useful integration for from-scratch lane-level HD map generation, and the paper is clearly written. The catch: the headline connectivity and endpoint-alignment numbers are computed after the pipeline's own A* reconnection, largest-component filtering, and lane geometry/topology repair passes, so they mostly show that the repair steps work. The 94.4% reduction in endpoint error versus baselines is apples-to-oranges because baselines don't get an equivalent repair pass.\n\nWhat is actually new: the coarse-to-fine pipeline--VQ-VAE/masked Transformer skeleton, structure-tensor procedural growth, and topology-aware lane construction--is a new system integration. The conditioning vector gives practical control over density and road-style attributes. Export to OSM/OpenDRIVE and the Tactics2D import path are real engineering contributions. Generation time of 1.39-3.50s is credible.\n\nSoft spots: no code or data; the growth-stage hyperparameters (sigma, w_tensor, w_inertia, theta_max, delta_s, snapping radius) and latent grid/codebook sizes are undisclosed, so the method is not reproducible as written. The validation in Section IV-E is import success and route planning only; no vehicle is driven. 'Simulation-ready' is a reasonable shorthand for 'imports into a simulator and supports routing,' but not for 'verified closed-loop behavior.' There are no ablations that show the contribution of each stage, and realism is only supported by qualitative figures. The stress-test note is correct on all these points.\n\nThese gaps are substantial but not fatal. The core system is plausible, and the external checks (100% import, 98.7% routing on 100 maps) give it a foundation. This paper deserves a serious referee. With code release, hyperparameters, ablations, and a closed-loop or traffic-simulation evaluation, it would be a strong reference for the autonomous-driving simulation community. As submitted, I would support conditional acceptance: require the code and data, and add at least one downstream driving evaluation.","headline":"RoadWeaver is a solid system integration for from-scratch HD map generation with practical deployability, but its headline metrics are largely by-construction and 'simulation-ready' is only shown via import and routing.","tokens_in":10223,"tokens_out":3245,"would_cite":true,"duration_ms":32602,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"RoadWeaver generates complete, large-scale, lane-level HD maps from scratch for autonomous-driving simulation, reporting 99.8% reachability, 0.24 m endpoint alignment, and direct import into a driving simulator.","keywords":["HD map generation","autonomous driving simulation","lane-level maps","road network generation","VQ-VAE","masked transformer","procedural road growth","closed-loop evaluation"],"falsifier":"Take a fixed rule-based lane-following and speed controller, run the same set of routes on 100 RoadWeaver maps and on 100 manually designed maps of comparable complexity, and count task completion, collisions, and off-road events. If completion collapses or collision rates are far higher on the generated maps, the claim that the maps are directly deployable for closed-loop evaluation fails.","tokens_in":9215,"feed_emoji":"🗺️","tokens_out":11504,"duration_ms":115143,"temperature":0.7,"pith_summary":"RoadWeaver aims to remove the map bottleneck in autonomous-driving simulation: instead of handcrafting maps or reconstructing real roads, it generates complete lane-level HD maps from an empty canvas. The paper's central claim is that a single coarse-to-fine pipeline can deliver global topological coherence, lane-level geometric validity, user control over density, and direct simulator compatibility at the same time. On the paper's metrics, the generated networks reach 99.8% reachability, keep only 10.7% dead ends, align expected lane endpoints to 0.24 m, and take 1.39–3.50 s to produce. A sympathetic reader would care because this would let evaluation suites scale to many diverse road networks on demand, which long-horizon closed-loop testing currently needs but handcrafted or reconstructed maps cannot easily supply.","feed_headline":"One pipeline weaves complete simulation-ready HD maps in seconds","feed_subtitle":"Generated road networks reach 99.8% connectivity and 0.24 m alignment, then import straight into a driving simulator.","key_machinery":"The central machinery is the road-field tensor $F \\in \\mathbb{R}^{H \\times W \\times 6}$, a six-channel raster carrying road probability, two orientation channels, junction and endpoint heatmaps, and a distance field, which lets a road graph be treated as an image-like field. A VQ-VAE compresses this tensor into discrete latent tokens, and a conditional masked Transformer, conditioned on an 11-dimensional vector of road-style and structural priors such as density, gridness, radialness, organicness, and bearing entropy, samples globally coherent skeletons. Expansion relies on the structure tensor $T(p)$ built from tangent samples along the skeleton; its eigenvectors steer a procedural road-growth front with style-dependent perturbation, and A* search over the predicted road evidence reconnects dangling ends. The lane-construction stage converts the refined graph into directed lanes and junction connectors, then runs geometry and topology repair passes, so the final map carries consistent predecessor, successor, and neighbour links.","core_discovery":"RoadWeaver's central claim is that from-scratch lane-level HD map generation can be decomposed into three stages that are individually tractable and jointly sufficient: sample a sparse global road skeleton from a learned discrete representation, expand it into a connected road graph by tensor-guided procedural growth, and convert that graph into a lane-level map with explicit topology-repair passes. The paper reports that this produces a directed lane graph with predecessor, successor, and neighbour relationships, a largest-connected-component ratio of 99.9%, reachability of 99.8%, a dead-end ratio of 10.7%, a cycle ratio of 85.2%, and a 0.24 m average distance between lane endpoints that should meet. It further reports that 100 generated maps all import into a driving simulator and that 987 of 1,000 route-planning tasks succeed, with full-map generation in 1.39–3.50 s. These numbers are offered as evidence that the pipeline closes the gap between generative map synthesis and practical closed-loop evaluation.","pith_inferences":["Because the paper's simulator test stops at route planning, a natural next experiment is a full closed-loop planner-controller evaluation; that experiment would show whether the 0.24 m endpoint alignment is sufficient for stable driving behavior.","The skeleton model and the procedural expansion are separable components, so a reader could replace the learned skeleton prior with a different city's distribution while keeping the growth rules, directly testing how much of the output's quality comes from the learned prior rather than the procedural geometry.","The endpoint-alignment error is measured on lane connections the pipeline itself computes and repairs; an independent audit measuring alignment on all near-miss lane endpoints before and after the repair pass would reveal whether the 0.24 m reflects raw generation quality or the repair step."],"forward_implications":["Simulation suites can be scaled to hundreds of diverse networks by increasing map area at fixed road density, removing the need to handcraft each environment.","A 0.24 m endpoint alignment error, compared with 4.32–8.17 m for the examined baselines, places generated lane geometry in the range where route planning in the tested simulator succeeds on 98.7% of tasks.","Because maps are exported in standard road-map formats, the same generated network can be moved across simulator toolchains without manual conversion.","Controllable road-node density gives evaluation designers a direct knob for sweeping map complexity while keeping connectivity and lane-level consistency."],"supporting_citations":[{"why":"Rule-based scenario-composition baseline that RoadWeaver compares against for topology quality and scalability.","marker":"[15]"},{"why":"Data-driven local HD-map generation baseline whose evaluation protocol for reachability, dead-end ratio, and endpoint alignment is adopted.","marker":"[16]"},{"why":"Open-source driving simulator used for deployment validation, with 100 generated maps imported and 1,000 route-planning tasks run.","marker":"[19]"},{"why":"Rule-based road-generation baseline whose geometric constraints motivate a scale-controllable growth strategy.","marker":"[23]"},{"why":"VQ-VAE discrete representation learning that RoadWeaver's road-field tokenization builds on.","marker":"[26]"},{"why":"Transformer-based discrete image synthesis that supports the VQ-VAE decoding setup used for road fields.","marker":"[27]"},{"why":"Masked token prediction method used to model the conditional distribution over latent road-field tokens.","marker":"[28]"},{"why":"Procedural city road-growth strategy that the skeleton-guided expansion stage follows.","marker":"[30]"},{"why":"Structure-tensor formulation used to build the directional tensor field guiding road expansion.","marker":"[31]"}],"fun_headline_variants":["RoadWeaver builds full HD maps from nothing in seconds","From zero to lane-level map: RoadWeaver does it in ~3s","HD map generation from scratch: 94% less error, 3s","Scratch-built HD maps for simulators: 99.8% reachable","RoadWeaver: complete HD maps, 1.4-3.5s, simulator-ready"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a map that scores well on the paper's own connectivity and alignment metrics, and that supports route planning in one simulator, will also support full closed-loop autonomous driving — the paper does not run any vehicle-level driving test.","fun_headline_variants_meta":{"raw":{"variants":["RoadWeaver builds full HD maps from nothing in seconds","From zero to lane-level map: RoadWeaver does it in ~3s","HD map generation from scratch: 94% less error, 3s","Scratch-built HD maps for simulators: 99.8% reachable","RoadWeaver: complete HD maps, 1.4-3.5s, simulator-ready"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000254,"raw_usage":{"total_tokens":1575,"prompt_tokens":962,"completion_tokens":613,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":508}},"tokens_in":578,"tokens_out":613,"duration_ms":6342,"temperature":1.0,"reasoning_tokens":508,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:34:17.929116+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed rule-based lane-following and speed controller, run the same set of routes on 100 RoadWeaver maps and on 100 manually designed maps of comparable complexity, and count task completion, collisions, and off-road events. If completion collapses or collision rates are far higher on the generated maps, the claim that the maps are directly deployable for closed-loop evaluation fails.","supporting_citations":[{"cited_title":"Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Rule-based scenario-composition baseline that RoadWeaver compares against for topology quality and scalability."},{"cited_title":"Hdmapgen: A hierarchical graph generative model of high definition maps,","cited_arxiv_id":null,"evidence_quote":"Data-driven local HD-map generation baseline whose evaluation protocol for reachability, dead-end ratio, and endpoint alignment is adopted."},{"cited_title":"Tactics2d: A highly modular and extensible simulator for driving decision-making,","cited_arxiv_id":null,"evidence_quote":"Open-source driving simulator used for deployment validation, with 100 generated maps imported and 1,000 route-planning tasks run."},{"cited_title":"Taming trans- formers for high-resolution image synthesis,","cited_arxiv_id":null,"evidence_quote":"Transformer-based discrete image synthesis that supports the VQ-VAE decoding setup used for road fields."},{"cited_title":"Maskgit: Masked generative image transformer,","cited_arxiv_id":null,"evidence_quote":"Masked token prediction method used to model the conditional distribution over latent road-field tokens."},{"cited_title":"Procedural modeling of cities,","cited_arxiv_id":null,"evidence_quote":"Procedural city road-growth strategy that the skeleton-guided expansion stage follows."},{"cited_title":"Optimal orientation detection of linear symmetry,","cited_arxiv_id":null,"evidence_quote":"Structure-tensor formulation used to build the directional tensor field guiding road expansion."}],"review_version":1}