Abstract
Validating autonomous vehicles (AVs) requires extensive simulation to ensure safety and reliability. However, existing scenario generation tools often demand significant programming expertise, creating a barrier for domain experts. This paper presents a no-code, interactive framework that leverages a graph-based map representation and an intuitive graphical user interface (GUI) to democratize scenario design in the CARLA simulator. By abstracting complex simulation logic into a visual editor, the framework enables rapid, high-fidelity testing environment creation without requiring manual coding.
Introduction
The validation of Autonomous Vehicles (AVs) is a monumental challenge. Estimates suggest that an AV would need to drive approximately 275 million incident-free miles to demonstrate it is just 20% safer than a human driver—a feat that is prohibitively expensive and dangerous in the real world. While simulation offers a viable alternative, current tools like Scenic or OpenSCENARIO often have a high barrier to entry, requiring specialized programming knowledge or deep expertise in simulation engines.
This technical gap excludes non-programmers, such as safety analysts and policymakers, from the validation process. The proposed framework aims to "democratize" scenario generation by providing a tool that combines the ease of a visual editor with the power needed for advanced deep learning research.
Methodology
The framework's core innovation lies in its four-stage pipeline and its underlying graph-based scenario representation.
Graph-Based Map Representation
The system ingests high-definition OpenDRIVE (.xodr) maps and converts them into a directed graph $\mathcal{G} = \{\mathcal{N}, \mathcal{E}\}$. In this model, nodes represent valid spawn points for actors, while edges capture critical road topology information, including successor/predecessor relationships and left/right lane adjacencies.
Interactive Scenario Design
- Region of Interest (ROI) Selection: Users interactively select "subgraphs" (e.g., specific intersections or highway segments) instead of loading entire resource-heavy maps. This allows for localized, high-intensity testing.
- Reachability Analysis: To define actor paths, users select spawn and goal nodes. The framework performs a reachability analysis to identify all valid terminal nodes, ensuring logically consistent paths through the network.
- No-Code GUI: A unified interface manages map selection, environmental configurations (weather, time of day), and actor placement. It includes a live bird's-eye view stream of the CARLA simulation.
- Automated Mode: For large-scale data generation, the framework can randomly sample parameters such as actor types, behaviors, and weather to create diverse datasets for stress-testing.
Data and Tools
The framework is integrated with the CARLA ecosystem, utilizing standard OpenDRIVE maps (e.g., Town01, Town04, Town10HD) and the CARLA asset catalog. By serializing extracted map data into lightweight JSON files, the framework supports offline scenario design, allowing users to configure environments without running the full simulator, thereby saving computational resources.
Results and Discussion
Qualitative evaluations demonstrate that scenarios designed via the GUI translate accurately into the CARLA environment, maintaining intended road layouts and agent logic. The framework successfully abstracts the complexity of CARLA’s Traffic Manager and Walker AI, allowing users to control nuanced behaviors—such as a cyclist riding slightly off-center—through simple parameter sliders.
Key Insights
- Graph vs. Raster: Unlike rasterized images, the sparse graph representation is lightweight and compatible with modern Graph Neural Networks (GNNs).
- Incremental ROI Expansion: The tool uses a strictly incremental process for defining testing areas, preserving the logical consistency of traffic flow.
Limitations and Future Work
Current limitations include a reliance on manual design or random sampling. Future work will focus on integrating real-world traffic data, incorporating heuristic-guided methods to target "edge cases," and developing automated metrics to score dataset diversity and challenge levels.
Summary updated