MARL-Based High-Risk Multi-Vehicle Scenario Generation for Autonomous Vehicle Safety Testing

Advancements in autonomous driving technology have been substantial, yet ensuring the safety and reliability of Autonomous Vehicles (AVs) in high-risk situations remains a critical challenge. Traditional mileage-based testing is often inadequate for capturing rare but safety-critical events, leading to a growing emphasis on scenario-based virtual simulation. This study proposes a novel method for generating high-risk multi-vehicle scenarios using Multi-Agent Reinforcement Learning (MARL) and in-depth crash data. A total of 567 real-world crash scenarios were extracted from the China In-depth Mobility Safety Study-Traffic Accident (CIMSS-TA) database and reconstructed to include both crash-involved and non-crash-involved vehicles that may have influenced the crash. A Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm is employed to model interactions among these vehicles and the AV under test, with custom-designed reward functions guiding adversarial agent behavior. The generated scenarios were validated using surrogate safety measures, assessing both risk levels and crash severity. Experimental results demonstrate that the proposed method produces significantly more complex and risk-intensive scenarios than baseline approaches, thereby exposing potential weaknesses in AV decision-making. Notably, the inclusion of non-crash-involved vehicles substantially elevates scenario difficulty, highlighting their critical role in realistic AV safety assessment. This framework offers a scalable approach to augmenting scenario diversity and enhancing the credibility of AV testing protocols.