CORTEX-AVD: CORner Case Testing&EXploration for Autonomous Vehicles Development

2025-04-04

AI Summary

Reliability verification of autonomous vehicles in rare, high-risk “corner cases” (e.g., sudden pedestrian intrusion, erratic lane changes) is hindered by prohibitive real-world data collection costs and challenges in accurate simulation modeling. To address this, we propose CORTEX-AVD—an open-source framework that uniquely integrates the CARLA simulator with the Scenic scenario description language, enabling automated generation of diverse traffic scenarios from natural-language specifications. We design a multi-factor fitness function incorporating distance, time, velocity, and collision probability, and employ genetic algorithms (GA) for efficient parameter optimization. Furthermore, we establish the first GA-driven benchmark for corner-case generation and evaluation. Experimental results across six representative scenario categories demonstrate that our approach significantly increases hazardous-event triggering rates while reducing无效 simulation iterations, achieving superior efficiency and scenario diversity compared to baseline methods.

Technology Category

• Search and Optimization: Sampling/Simulation-based Search
• Planning, Routing, and Scheduling: Planning with Language Models
• Multiagent Systems: Adversarial Agents

Application Category

• Web Mining and Content Analysis: Web data generation and simulation
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: User modeling and simulation for interactive and conversational systems

Abstract

Autonomous Vehicles (AVs) aim to improve traffic safety and efficiency by reducing human error. However, ensuring AVs reliability and safety is a challenging task when rare, high-risk traffic scenarios are considered. These 'Corner Cases' (CC) scenarios, such as unexpected vehicle maneuvers or sudden pedestrian crossings, must be safely and reliable dealt by AVs during their operations. But they arehard to be efficiently generated. Traditional CC generation relies on costly and risky real-world data acquisition, limiting scalability, and slowing research and development progress. Simulation-based techniques also face challenges, as modeling diverse scenarios and capturing all possible CCs is complex and time-consuming. To address these limitations in CC generation, this research introduces CORTEX-AVD, CORner Case Testing&EXploration for Autonomous Vehicles Development, an open-source framework that integrates the CARLA Simulator and Scenic to automatically generate CC from textual descriptions, increasing the diversity and automation of scenario modeling. Genetic Algorithms (GA) are used to optimize the scenario parameters in six case study scenarios, increasing the occurrence of high-risk events. Unlike previous methods, CORTEX-AVD incorporates a multi-factor fitness function that considers variables such as distance, time, speed, and collision likelihood. Additionally, the study provides a benchmark for comparing GA-based CC generation methods, contributing to a more standardized evaluation of synthetic data generation and scenario assessment. Experimental results demonstrate that the CORTEX-AVD framework significantly increases CC incidence while reducing the proportion of wasted simulations.

Problem

• Generates rare high-risk scenarios for AV testing
• Automates corner case creation via text descriptions
• Optimizes scenarios using multi-factor genetic algorithms

Innovation

• Integrates CARLA and Scenic for automated CC generation
• Uses Genetic Algorithms to optimize scenario parameters
• Multi-factor fitness function enhances risk assessment