115. ProSafeAV: Enhancing Proactive Safety Performance of AVs using Wo…

International Oral Session, International, 27(금) 13:20-15:00, 라카이볼룸Ⅰ 

ProSafeAV: Enhancing Proactive Safety Performance of AVs using World Model-Based Reinforcement Learning and Extreme Value Theory / Yukai Wang(Korea Advanced Institute of Science and Technology), Tiantian Chen(Korea Advanced Institute of Science and Technology), Sikai Chen(University of Wisconsin-Madison), Kitae Jang(Korea Advanced Institute of Science and Technology)

The increasing deployment of autonomous vehicles (AVs) necessitates robust safety mechanisms to navigate complex traffic scenarios. This study introduces ProSafeAV, a novel framework integrating Extreme Value Theory (EVT) with world model-based reinforcement learning (RL) to enhance AVs' proactive safety and performance. Traditional approaches often rely on reactive safety measures, leaving AVs vulnerable to traffic conflicts. ProSafeAV addresses this limitation by emphasizing prediction and prevention of high-risk traffic situations. Our methodology integrates EVT with world model-based RL in a comprehensive architecture. The system processes environmental data through sensor fusion and safety indicators to predict and avoid potential conflicts. Key components include a perception module for sensor data fusion, a world model for efficient latent space representation, and an RL module for decision-making. The EVT module analyzes tail behavior of conflict-related metrics, enhancing risk assessment. We evaluated ProSafeAV using the CARLA simulation platform and CarDreamer model, focusing on challenging scenarios such as overtaking tasks. Our experiments compared ProSafeAV against benchmark models DreamerV2 and DreamerV3. Results demonstrate ProSafeAV's significant outperformance in key metrics including average Time-to-Collision (TTC), collision avoidance, and overall driving score. ProSafeAV achieved higher TTCs while maintaining competitive speeds, indicating a balanced approach to safety and efficiency. The integration of EVT with RL frameworks enables enhanced risk assessment and decision-making, allowing AVs to anticipate and mitigate potential conflicts effectively. These findings suggest that ProSafeAV's approach could significantly advance AV safety and reliability in dynamic, complex environments.