RoadWeaver: Large-Scale Lane-Level HD Map Generation from Scratch for Autonomous Driving Simulation

RoadWeaver is a coarse-to-fine framework for from-scratch generation of diverse, large-scale lane-level HD maps for autonomous driving simulation. It addresses limitations of existing approaches that either rely on handcrafted or reconstructed real-world maps, which limits scalability, or generate only local road structures rather than complete HD maps.

The framework comprises three stages: global road skeleton generation, skeleton-guided road graph expansion, and topology-aware HD map construction. The first stage generates a sparse road skeleton that captures the principal structure of the target road network, using a VQ-VAE to encode a road field tensor into discrete latent tokens and a conditional masked Transformer to model the distribution of these tokens conditioned on a vector containing a six-dimensional road-style code together with five structural priors: road density, gridness, radialness, organicness, and bearing entropy. The second stage expands the skeleton using a procedural road-growth strategy based on a structure tensor field, with dangling endpoints reconnected via A* search and residual disconnected fragments removed through largest-connected-component filtering. The final stage transforms the refined road graph into a lane-level HD map through a sequence of graph transformation and lane construction operations, including lane configuration assignment, lane boundary generation, junction connector construction, and topology repair.

Experimental results show RoadWeaver achieves a 99.8% reachability, a 10.7% dead-end ratio, and an endpoint alignment error of 0.24 m. Compared with state-of-the-art generation methods (MetaDrive, RoadGen, HDMapGen), it reduces endpoint alignment error by 94.4% while generating complete HD maps in 1.39–3.50 s. The cycle ratio reaches 85.2%, compared with 50.5% for MetaDrive, 46.5% for HDMapGen, and 0.0% for RoadGen, indicating more closed-loop structures, resulting in more connected road blocks and multiple routing possibilities.

RoadWeaver supports controllability over road-network density: By increasing the target density, the generated maps become progressively denser while maintaining connectivity and lane-level consistency, with measured densities remaining close to specified values. It scales by enlarging the map area while maintaining a bounded road density, so the number of generated nodes can grow naturally without changing the generation procedure.

The generated maps are simulation-ready: The generated maps are compatible with OSM and OpenDRIVE representations and can be imported into Tactics2D without manual conversion. In validation, All generated maps are successfully imported into the simulator, achieving an import success rate of 100.0%, and The built-in route planner successfully generates valid paths for most of the tasks, corresponding to a routing success rate of 98.7%.

Improvements to AI systems:

1. Hierarchical generative planning with discrete latent tokens: Implement a two-level generative model where a global sparse structure (skeleton) is first generated via a VQ-VAE + masked Transformer, then expanded locally. This improves scalability and coherence for large-scale outputs (e.g., city layouts, molecule graphs, or circuit designs) compared to single-shot generation.

2. Controllable structural priors via vector conditioning: Add explicit, interpretable control knobs (e.g., density, gridness, organicness, bearing entropy) as conditioning inputs to the generative model. This allows users to steer output characteristics (e.g., road network style, urban vs. rural) without retraining, enabling fine-grained, multi-objective generation.

3. Procedural expansion with graph repair heuristics: Use a structure-tensor-guided growth algorithm for local expansion, followed by A* search to reconnect dangling endpoints and largest-connected-component filtering to remove fragments. This improves connectivity and reduces dead-ends in any graph-based generative output (e.g., road networks, vascular systems, or utility grids).

4. Topology-aware post-processing for simulation readiness: Integrate a final stage that converts the raw graph into a standardized, simulation-compatible format (e.g., lane-level attributes, junction connectors, topology repair). This ensures the generated output is directly usable by downstream simulators or planners without manual conversion, improving end-to-end automation.

5. Closed-loop structure optimization: Explicitly optimize for high cycle ratio (closed loops) during generation, which improves redundancy and multiple routing possibilities. This can be applied to any network generation task (e.g., communication networks, logistics routes) to enhance resilience and alternative path availability.

6. Scalable generation with bounded density: Design the generation procedure to scale with area size while keeping node density bounded, so the system can produce arbitrarily large outputs without changing the algorithm. This enables generation of city-scale or continent-scale structures with consistent local quality.

---

What the improved AI system can do:

• Generate large-scale, coherent, and controllable spatial networks (roads, pipelines, power grids) from scratch in seconds, with user-specified density and style.

• Produce outputs that are immediately usable in simulation environments (e.g., autonomous driving, disaster response) with 100% import success and >98% routing success.

• Maintain high connectivity (99.8% reachability) and low dead-end ratios (10.7%) even at scale, while offering multiple alternative paths (85% cycle ratio).

• Allow users to adjust generation parameters (e.g., density) on the fly, observing progressive densification without breaking lane-level consistency.

• Automatically repair topological errors (dangling ends, disconnected fragments) during generation, eliminating manual post-processing.

• Scale from a small neighborhood to a full city or region without retraining, by simply enlarging the target area and keeping density bounded.