A Trend-Guided Testing Framework Balancing Risk and Coverage to Accelerate Autonomous Vehicle Safety Validation
Authors:
- Jiang BianJBJiang Bian
- Zimo YangZYZimo Yang
- Gui GuiGGGui Gui
- Rui ZhouRZRui Zhou
- Helai HuangHHHelai Huang
Preprint
- 19 May 2026
- preprint
- Published by Elsevier in SSRN Electronic Journal
Abstract
Scenario-based testing has become the mainstream approach in Autonomous Vehicle (AV) safety validation, with increasingly diverse methods for constructing scenario libraries. Although scenario generation primarily aims to produce high-risk scenarios, exhaustively testing all scenarios remains resource-intensive. To address this challenge, this study proposes an accelerated testing framework, the Trend-Guided Testing Framework (TGTF), to enhance AV safety validation. The framework analyzes tested scenario distributions and leverages observed crash trends to predict the distribution of high-risk, untested scenarios. This prediction then guides the selection of scenarios for the next round of simulation testing. Since the trend analysis of scenario distribution is conducted in a low-dimensional space, while scenarios themselves exist in a high-dimensional space, TGTF incorporates a Genetic Algorithm (GA)-based high-dimensional information reconstruction method. Furthermore, a high-performance scenario generation model is integrated into the framework. Experimental validation was conducted using Baidu Apollo as the AV control system. The results indicate that the selected simulation scenarios were not only diverse but also highly effective in identifying critical safety risks. For the remaining scenarios, surrogate safety metrics such as PET and GTTC confirmed that they still exhibited high-risk characteristics. These findings, along with several ablation studies and comparative studies against two baselines, highlight the effectiveness of TGTF in prioritizing high-risk scenarios while maintaining testing coverage and optimizing resource allocation in AV safety testing.Keywords
This publication has 40 references indexed in Scilit:
- Search-to-Crash: Generating safety-critical scenarios from in-depth crash data for testing autonomous vehiclesEnergy, 2025
- Quantile-based scenario generation for automated vehicle safety evaluationAccident Analysis & Prevention, 2025
- How would autonomous vehicles behave in real-world crash scenarios?Accident Analysis & Prevention, 2024
- A Survey on an Emerging Safety Challenge for Autonomous Vehicles: Safety of the Intended FunctionalityEngineering, 2024
- Poisson-Disk Sampling: Theory and Applications2024
- Toward Trustworthy Decision-Making for Autonomous Vehicles: A Robust Reinforcement Learning Approach with Safety GuaranteesEngineering, 2023
- Key Challenges and Chinese Solutions for SOTIF in Intelligent Connected VehiclesEngineering, 2023
- Towards the Unified Principles for Level 5 Autonomous VehiclesEngineering, 2021
- Instance spaces for machine learning classificationMachine Learning, 2017
- Towards objective measures of algorithm performance across instance spaceComputers & Operations Research, 2014