Andrii Bidochko

• Updated: August 20, 2026
• 7 min read

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

Direct Answer

RoadWeaver is a coarse‑to‑fine framework that can generate complete, lane‑level high‑definition (HD) maps from scratch, scaling to city‑size road networks in a few seconds. This capability removes the bottleneck of manually crafting or reconstructing maps, enabling autonomous‑driving simulators to run large‑scale, closed‑loop evaluations without the cost and latency of real‑world data collection.

Background: Why This Problem Is Hard

Autonomous‑vehicle (AV) developers rely on HD maps to provide precise lane geometry, traffic‑sign semantics, and connectivity information that sensors alone cannot guarantee. However, creating such maps faces three intertwined challenges:

• Scalability. Hand‑crafted maps require teams of cartographers and can take weeks to cover a single urban district.
• Real‑world reconstruction limits. Existing map‑reconstruction pipelines depend on LiDAR or aerial imagery, which are expensive, weather‑sensitive, and often incomplete in dense urban canyons.
• Simulation fidelity. Current procedural generators produce only local road fragments, forcing developers to stitch together dozens of pieces manually, which leads to topological inconsistencies such as dead‑ends or misaligned lane connections.

Because simulation is the primary safety net before road testing, any gap in map realism directly translates into riskier deployments. The industry therefore needs a method that can produce large, topologically correct, lane‑level maps quickly and without external data dependencies.

What the Researchers Propose

The authors introduce RoadWeaver, a three‑stage pipeline that synthesizes a global road layout, expands it into a fully connected network, and finally refines each road segment into lane‑level geometry with consistent connectivity. The framework is built around two core ideas:

1. Coarse‑to‑fine synthesis. A high‑level graph representing road corridors is first generated using a learned generative model. This graph captures macro‑scale topology (e.g., intersections, highway spines) without committing to precise lane shapes.
2. Topological consistency enforcement. As the graph is densified into lane geometry, a set of constraints guarantees that every lane has a valid predecessor and successor, eliminating dead‑ends and ensuring that lane endpoints align within centimeters.

Key components include:

• Global Layout Generator – a diffusion‑based model that outputs a sparse road skeleton conditioned on user‑specified city size and road density.
• Network Expansion Module – a rule‑based engine that inserts missing connections, resolves intersections, and guarantees a single‑connected graph.
• Lane‑Level Builder – a neural‑geometric network that converts each edge of the expanded graph into multi‑lane polylines, applying curvature smoothing and lane‑width heuristics.
• Consistency Checker – a post‑processing validator that detects and corrects endpoint misalignments, dangling lanes, and non‑manifold junctions.

How It Works in Practice

Step‑by‑Step Workflow

The RoadWeaver pipeline can be visualized as a linear flow, but each stage interacts with the others through feedback loops:

1. Input Specification. Users provide high‑level constraints such as target area (e.g., 10 km²), desired road density, and optional seed points for major arteries.
2. Coarse Graph Generation. The Global Layout Generator samples a road‑skeleton graph that respects the constraints. This graph is a set of nodes (junctions) and edges (road segments) without lane detail.
3. Connectivity Augmentation. The Network Expansion Module examines the skeleton for isolated components, adds missing edges, and resolves ambiguous intersections using a deterministic rule set.
4. Lane Geometry Synthesis. For each edge, the Lane‑Level Builder predicts the number of lanes, lane widths, and curvature profiles. The output is a set of parallel polylines representing individual lanes.
5. Topological Validation. The Consistency Checker runs a graph‑theoretic analysis to ensure every lane endpoint matches a counterpart on an adjacent lane. Misalignments are corrected by a local optimization that nudges points within a 0.25 m tolerance.
6. Export. The final product is an HD map in standard OpenDRIVE or Lanelet2 format, ready for immediate ingestion by simulators such as CARLA, LGSVL, or proprietary platforms.

What Sets RoadWeaver Apart

Unlike prior procedural generators that stop at the road‑network level, RoadWeaver delivers lane‑level detail while preserving global topology. Its diffusion‑based coarse generator provides diversity comparable to real‑world city layouts, yet the deterministic expansion guarantees a single, navigable graph—something purely stochastic methods struggle to achieve. Moreover, the entire pipeline runs in under four seconds on a single GPU, making it practical for on‑the‑fly scenario generation during simulation campaigns.

Evaluation & Results

The authors benchmarked RoadWeaver against two state‑of‑the‑art map generators and a baseline reconstruction pipeline. Evaluation focused on three dimensions:

• Reachability. The proportion of lanes that can be traversed from any start point to any destination within the generated map.
• Dead‑End Ratio. The fraction of lane endpoints that do not connect to another lane, indicating topological errors.
• Endpoint Alignment Error. The average Euclidean distance between paired lane endpoints at intersections.

Key findings include:

• RoadWeaver achieved a 99.8 % reachability score, meaning virtually every lane participates in a complete route network.
• The dead‑end ratio dropped to 10.7 %, a ten‑fold improvement over the nearest competitor.
• Endpoint alignment error was reduced to 0.24 m, a 94.4 % decrease compared with the previous best method.
• Generation time ranged from 1.39 s to 3.50 s for maps covering up to 25 km², far faster than reconstruction pipelines that require minutes to hours.

Beyond raw metrics, the authors demonstrated that the generated maps could be dropped directly into the CARLA simulator, where autonomous agents completed multi‑kilometer routes without encountering navigation failures. This end‑to‑end validation underscores the practical readiness of RoadWeaver for large‑scale simulation pipelines.

Why This Matters for AI Systems and Agents

Simulation is the crucible where perception, planning, and control modules are stress‑tested before real‑world deployment. RoadWeaver reshapes that crucible in three concrete ways:

• Accelerated scenario generation. Engineers can spin up new city‑scale maps on demand, enabling rapid A/B testing of perception stacks across diverse road topologies.
• Consistent evaluation metrics. Because the generated maps guarantee reachability and low dead‑end rates, performance differences observed across runs can be attributed to algorithmic changes rather than map artifacts.
• Cost reduction. Eliminating the need for expensive LiDAR surveys or manual cartography translates into millions of dollars saved for OEMs and simulation providers.

For AI agents that rely on high‑definition map data—such as route planners, behavior prediction models, and reinforcement‑learning policies—having a reliable, programmable source of lane‑level maps opens new research avenues. Teams can now train agents on procedurally varied environments, improving generalization to unseen cities.

Practically, developers can integrate RoadWeaver‑generated maps into existing pipelines using the UBOS platform overview for data orchestration, or leverage the Workflow automation studio to trigger map generation as part of a CI/CD workflow for autonomous‑driving software.

What Comes Next

While RoadWeaver marks a significant leap, several open challenges remain:

• Semantic enrichment. Current outputs focus on geometry; adding traffic‑sign, signal, and surface‑type annotations would further close the simulation‑to‑reality gap.
• Dynamic environment coupling. Integrating traffic‑flow generators that respect the lane topology could produce more realistic congestion patterns.
• Cross‑domain adaptation. Extending the diffusion model to respect regional design standards (e.g., European roundabouts vs. US grid layouts) would broaden applicability.

Future research may explore multimodal conditioning—feeding satellite imagery or city planning data into the coarse generator—to produce maps that are both procedurally diverse and geographically plausible. Additionally, coupling RoadWeaver with a OpenAI ChatGPT integration could enable natural‑language specifications of map constraints, making the tool accessible to non‑technical stakeholders.

From a product perspective, the authors plan to release an open‑source SDK that plugs directly into popular simulation environments. This will allow developers to embed map generation into their test harnesses, turning map creation from a pre‑step into an on‑the‑fly capability.

References

• RoadWeaver: Large‑Scale Lane‑Level HD Map Generation from Scratch for Autonomous Driving Simulation – Original arXiv preprint, August 2026.
• CARLA Simulator – Open-source autonomous‑driving simulation platform.
• OpenDRIVE Specification – Standard format for HD map exchange.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.