Source-linked AI summary
RoadWeaver: Large-Scale Lane-Level HD Map Generation from Scratch for Autonomous Driving Simulation
Yueyuan Li, Zexi Chen, Weijie Xi, Mingyang Jiang, Songan Zhang, Hanyang Zhuang, Ming Yang
TL;DR
Long-horizon autonomous-driving simulation needs diverse, scalable maps with lane-level detail, while existing approaches are limited in scalability or generate only partial map structures. RoadWeaver generates large-scale lane-level HD maps from scratch while preserving topology and geometry. It reports 99.8% reachability and a 0.24 m endpoint alignment error, the smallest among compared baselines.
Problem
Long-horizon evaluation requires diverse, large-scale road maps with lane-level geometry and connectivity, but existing map approaches have scalability limits or struggle to jointly provide global coherence, valid geometry, controllability, and simulator compatibility.
Source: §I · p. 1Method
RoadWeaver generates diverse, large-scale lane-level HD maps from scratch while preserving road-network topology and lane geometry.
Source: §I · p. 1Results
99.8% reachability and 0.24 m endpoint alignment error, the smallest among compared methods whose errors range from 4.32 m to 8.17 m.
Source: C. Topological and Geometric Validity · p. 6Takeaways & Limitations
Generated maps are diverse, scalable, and compatible with existing driving simulators, making them suitable for large-scale closed-loop ADS evaluation.
Source: §V · p. 7
Abstract
from arXiv · showhide
Autonomous driving simulation requires diverse and scalable lane-level HD maps to support long-horizon evaluation across complex road networks. Existing approaches either rely on handcrafted or reconstructed real-world maps, which limits scalability, or generate only local road structures rather than complete HD maps. We present RoadWeaver, a coarse-to-fine framework for from-scratch generation of diverse, large-scale HD maps. RoadWeaver first synthesizes a global road layout, expands it into a connected road network, and then constructs lane-level geometry with topologically consistent lane connectivity. Experimental results show that RoadWeaver achieves a 99.8\% reachability, a 10.7\% dead-end ratio, and an endpoint alignment error of 0.24 m. Compared with SOTA generation methods, it reduces endpoint alignment error by 94.4\% while generating complete HD maps in 1.39--3.50 s. The generated maps can be directly deployed in driving simulators, providing scalable simulation environments for future closed-loop evaluation of autonomous driving systems. The training code and an out-of-the-box implementation of RoadWeaver will be released upon acceptance.
I. INTRODUCTION
Long-horizon closed-loop evaluation needs large, diverse maps with lane-level detail, while existing map-generation approaches struggle to provide global coherence, valid geometry, controllable morphology, and simulator compatibility together. RoadWeaver generates lane-level maps from scratch and supports direct use in simulation.
I. INTRODUCTION
- Long-horizon evaluation requires large, geographically diverse road maps with lane-level geometry and connectivity.Such evaluations can expose accumulated planning failures, including inappropriate lane selections, route deviations, and traffic-rule violations.Source: §I · p. 1
- Road topology influences driving and interaction patterns, making topological diversity important for broadening behavioral coverage.The paper reports distinct trajectory distributions across road topologies and analyzes velocity, acceleration, steering, and other motion characteristics.Source: §I · p. 1, Figure 1 · p. 1
- Handcrafted and reconstructed maps have scalability constraints, while automatic methods struggle to jointly provide global coherence, valid lane geometry, controllable morphology, and simulator compatibility.Handcrafted maps require substantial engineering effort; reconstructed maps are restricted to existing regions and their inherent topology.Source: §I · p. 1
- RoadWeaver generates diverse, large-scale lane-level HD maps from scratch while preserving road-network topology and lane geometry.Maps can be exported to OSM and OpenDRIVE and integrated with Tactics2D for customized long-horizon ADS evaluation.Source: §I · pp. 1–2
II. RELATED WORKS
Prior map-preparation work spans handcrafted construction, reconstruction from real-world observations, and automatic generation. Existing approaches address different aspects, but none is described as jointly supporting large-scale, simulation-ready HD maps for comprehensive ADS evaluation.
II. RELATED WORKS
- Handcrafted maps support controlled, reproducible experiments, but scaling them requires substantial engineering effort and their structural diversity is limited by available assets and manual design.CARLA and Highway-env are given as examples of simulators with handcrafted road environments.Source: §II · p. 2
- Real-world reconstruction methods can achieve high geometric fidelity but remain restricted to existing road networks, making novel topologies and large-scale maps costly to generate.Learning-based examples include HDMapNet, MapTR [21], and VectorMapNet.Source: §II · p. 2
- Rule-based generators offer efficiency and direct attribute control, but depend on manually specified primitives and rules that constrain structural diversity and lane-geometry variation.MetaDrive and RoadGen [23] are examples; scaling or increasing topology diversity often requires more templates and rules.Source: §II · p. 2
- Data-driven methods can learn richer road patterns, but existing examples focus on local synthesis or bounded patches rather than complete, simulation-ready large-scale lane maps with consistent connectivity.The paper identifies HDMapGen and ControlMap [24] as examples.Source: §II · p. 2
- Existing methods improve different aspects of map generation, but none simultaneously supports large-scale, simulation-ready HD maps for comprehensive ADS evaluation.The paper groups map preparation into manual construction, real-world reconstruction, and automatic generation, with automatic methods further divided into rule-based and data-driven approaches.Source: §II · p. 2
III. METHOD
RoadWeaver uses a coarse-to-fine pipeline, moving from a global road skeleton to a refined road graph and then a lane-level HD map. Its skeleton-generation stage represents road structure as a field and models that field through conditioned discrete latent tokens.
A. Overview
- RoadWeaver constructs maps in three stages: global road skeleton generation, skeleton-guided road graph expansion, and topology-aware HD map construction.The skeleton captures the principal network structure; branch and local roads add detail before conversion to an HD map with valid connectivity and topology.Source: A. Overview · p. 2
- Map scale is controlled by changing spatial extent while maintaining target road-network density, increasing node count without changing the generation procedure.This supports generating larger networks while retaining consistent structural characteristics.Source: A. Overview · p. 2
B. Global Road Skeleton Generation
- The initial road skeleton preserves long-range connectivity and overall layout while omitting geometric detail, allowing local roads to be added later without substantially changing global organization.This decomposition reduces the complexity of the initial generation task.Source: B. Global Road Skeleton Generation · p. 3
- Road graphs are rasterized into a six-channel road-field tensor encoding road probability, orientation, junctions, endpoints, and distance for subsequent graph reconstruction.During training, the field is constructed from ground-truth road graphs extracted from OSM.Source: B. Global Road Skeleton Generation · p. 3
- A VQ-VAE encodes the road field as discrete latent tokens, retaining dominant topology while suppressing geometric detail.The model is optimized with reconstruction, codebook, and commitment losses.Source: B. Global Road Skeleton Generation · p. 3
- A conditional masked Transformer models latent tokens using a vector of road-style and structural priors, with AdaLN injecting the conditioning into each block.The priors include road density, gridness, radialness, organicness, and bearing entropy; masked prediction uses bidirectional context to refine regions in parallel.Source: B. Global Road Skeleton Generation · p. 3
- At inference, the decoded road field yields centerlines that are extracted and reconstructed as a vectorized road skeleton.This completes the skeleton-generation stage before subsequent road-graph expansion.Source: B. Global Road Skeleton Generation · p. 3
C. Skeleton-Guided Road Graph Expansion
RoadWeaver expands a sparse skeleton into a road graph using directional guidance from a tensor field built from the skeleton. It reconnects and supplements roads to produce a more complete network.
C. Skeleton-Guided Road Graph Expansion
- RoadWeaver applies procedural road growth to add local detail while retaining the skeleton’s principal connectivity and large-scale organization.The skeleton intentionally omits geometric detail so later local-road additions do not substantially change the global network organization.Source: C. Skeleton-Guided Road Graph Expansion · p. 3
- Growth selects the tensor eigen-direction most aligned with the current front, combines tensor guidance with directional inertia, and limits turning by θ_max.The major and minor eigenvectors provide candidate directions; scalar weights control tensor guidance and inertia.Source: C. Skeleton-Guided Road Graph Expansion · p. 4
- The selected direction receives a style-dependent angular perturbation, and the growth front advances by a first-order Euler step of length Δs.Growth stops at the map boundary, a prescribed length or iteration limit, or when it approaches an existing road within the snapping radius; nearby roads are then connected.Source: C. Skeleton-Guided Road Graph Expansion · p. 4
- Dangling endpoints are reconnected with A* paths guided by the decoded road field, then disconnected fragments are removed by largest-connected-component filtering.The field assigns lower traversal costs to regions with stronger predicted road evidence.Source: C. Skeleton-Guided Road Graph Expansion · p. 4
- A continuous directional tensor field built from skeleton tangents guides growth into large unoccupied regions, helping local roads follow the surrounding network.The tensor combines nearby tangent directions with spatial contributions controlled by a Gaussian-kernel bandwidth.Source: C. Skeleton-Guided Road Graph Expansion · pp. 3–4
D. Lane-Level HD Map Construction
RoadWeaver turns a semantic road graph into a lane-level HD map through graph refinement, lane construction, and topology repair. The process establishes a directed lane graph and repairs geometric conflicts and disconnected connections.
D. Lane-Level HD Map Construction
- The final pipeline refines the road graph, assigns lane configurations, generates lane boundaries and directional lanes, and constructs junction connector lanes.It merges lane instances, adds successor links, repairs lane geometry, assembles the HD map, and applies a final topology repair pass.Source: D. Lane-Level HD Map Construction · p. 4
- Lane-level topology is established as a directed lane graph, with a final repair pass intended to recover disconnected lane connections.The pipeline also repairs geometric conflicts before assembling the map.Source: D. Lane-Level HD Map Construction · p. 4
- The pipeline takes a semantic road graph with geometry and road attributes as input and produces an HD map containing lane geometries and topology.Source: D. Lane-Level HD Map Construction · p. 4
1) Dataset:
RoadWeaver is evaluated on OSM-derived road-network data and compared with rule-based and data-driven map-generation baselines. The reported examples show coherent lane-level structure, variation at a fixed density, and density control that weakens above approximately 40 road nodes/km².
1) Dataset:
- Training uses approximately 58,000 OSM road-network samples from 144 cities, with each sample covering about 2 km × 2 km.Road-node density is defined as nodes per km² and is used as a controllable density measure; map area can be adjusted while maintaining a target density.Source: 1) Dataset · p. 4
2) Baselines:
- The comparison includes MetaDrive and RoadGen as rule-based baselines and HDMapGen as a data-driven baseline for local lane-level map generation.Source: 2) Baselines · p. 4
3) Hardware:
- Generation time is measured over multiple independent runs on an RTX 5090 workstation, excluding model loading and warm-up.The setup also uses an AMD Ryzen 7 9800X3D CPU and 64 GB of system RAM.Source: 3) Hardware · p. 4
B. Generation Quality, Diversity, and Controllability
- RoadWeaver produces coherent topology and geometry across map regions, maintaining road-segment connectivity and lane-level structures where baselines show disconnected components or abrupt transitions.The visual comparison uses methods configured for approximately 35–40 graph nodes; intersections, endpoints, and turning transition points count as nodes.Source: B. Generation Quality, Diversity, and… · pp. 4–5, Figure 3 · p. 5
- Maps generated with different seeds show distinct topology and geometry at the same target density, while increasing target density produces denser connected maps with lane-level consistency.Measured densities remain close to specified values, but density increases become less significant above approximately 40 road nodes/km²; Fig. 4 presents 2 km × 2 km maps across target densities of 10–40 road nodes/km².Source: B. Generation Quality, Diversity, and… · p. 5, Figure 4 · p. 5
C. Topological and Geometric Validity
RoadWeaver produces highly connected road networks with lane-level geometry that remains well aligned. Its generated graphs contain more closed loops than the compared baselines, while its endpoint alignment error is the lowest reported.
C. Topological and Geometric Validity
- 99.9% LCC and 99.8% reachability indicate that most generated road segments belong to a connected network, and RoadWeaver has the lowest dead-end ratio among compared methods.The evaluation uses graph-connectivity and network-structure metrics alongside a lane-level geometric metric.Source: C. Topological and Geometric Validity · pp. 5–6, Table II · p. 6
- 85.2% cycle ratio for RoadWeaver compares with 50.5% for MetaDrive, 46.5% for HDMapGen, and 0.0% for RoadGen.The paper reports that the larger cycle ratio corresponds to more closed-loop structures, connected road blocks, and routing possibilities.Source: C. Topological and Geometric Validity · p. 6
- 0.24 m endpoint alignment error is the smallest reported, compared with baseline values ranging from 4.32 m to 8.17 m.The metric is the average Euclidean distance between lane endpoints expected to connect.Source: C. Topological and Geometric Validity · p. 6
- RoadWeaver generates large-scale HD maps from scratch while maintaining graph connectivity and lane-level geometric consistency.The evaluation assesses connectivity, network structure, and endpoint alignment.Source: C. Topological and Geometric Validity · pp. 5–6
D. Scalability and Efficiency
RoadWeaver scales by enlarging map area while maintaining bounded road density, without changing its generation procedure or relying on fixed templates. It generates maps in approximately 1.39–3.50 s, while the compared methods face scale-related limitations.
D. Scalability and Efficiency
- RoadWeaver scales by enlarging map area at bounded road density, allowing node counts to grow without changing the generation procedure or using a fixed map template.The methods differ in how they scale: MetaDrive increases a predefined configuration, while RoadGen may terminate when its connection queue is exhausted and rarely exceeds roughly 60 nodes.Source: D. Scalability and Efficiency · p. 6
- Approximately 50 s to 370 s is RoadGen’s generation-time increase, making large-scale expansion impractical.HDMapGen can increase graph size, but larger graphs provide limited quality improvement.Source: D. Scalability and Efficiency · p. 6
- 1.39–3.50 s is the reported RoadWeaver map-generation time, with runtime comparable to MetaDrive and better scalability and road-network quality.Figure 5 compares generation time and CPU consumption as graph-node count increases.Source: D. Scalability and Efficiency · p. 6, Figure 5 · p. 6
E. Simulator Deployment Validation
RoadWeaver maps were imported into Tactics2D without manual conversion, editing, or postprocessing, and most sampled routing tasks succeeded. The conclusion positions the framework as a source of scalable, simulator-compatible maps for closed-loop ADS evaluation, while identifying richer traffic semantics, roadside assets, and interactive editing as future work.
E. Simulator Deployment Validation
- RoadWeaver maps are compatible with OSM and OpenDRIVE and can be imported into Tactics2D without manual conversion.Tactics2D is used to verify direct deployment for autonomous-driving simulation.Source: E. Simulator Deployment Validation · p. 6
- 100.0% of 100 generated maps imported into Tactics2D, and valid paths were generated for 98.7% of 1,000 routing tasks.Average routing time was about 0.82 ± 0.17 s; maps were imported without manual editing or postprocessing.Source: E. Simulator Deployment Validation · p. 6
V. CONCLUSION
- RoadWeaver’s generated maps are described as suitable for large-scale closed-loop ADS evaluation; future work targets richer traffic semantics, roadside assets, and interactive map editing.The conclusion characterizes the maps as diverse, scalable, and directly compatible with existing driving simulators.Source: §V · p. 7