Monte Carlo Tree Search for Behavior Planning in Autonomous Driving

• Published in: IEEE International Symposium on Safety, Security and Rescue Robotics pp. 117 - 124
• Main Authors: Wen, Qianfeng, Gong, Zhongyi, Zhou, Lifeng, Zhang, Zhongshun
• Format: Conference Proceeding
• Language: English
• Published: IEEE 12.11.2024
• Subjects: Autonomous vehicles; Decision making; Heuristic algorithms; Monte Carlo methods; Navigation; Planning; Safety; Search problems; Security; Vehicle dynamics
• ISSN: 2475-8426
• DOI: 10.1109/SSRR62954.2024.10770028

The integration of autonomous vehicles into urban and highway environments necessitates the development of robust and adaptable systems for behavior planning and decision-making. This study introduces a Monte Carlo Tree Search (MCTS)-based algorithm designed to navigate the challenges of autonomous driving. Our core objective is to exploit the balance between exploration and exploitation offered by MCTS to enable intelligent driving decisions in complex scenarios. We present a Monte Carlo Tree Search (MCTS)-based algorithm specifically designed for the nuanced requirements of autonomous driving. This approach incorporates meticulously devised cost functions that address safety, comfort, and efficiency, seamlessly integrated into the MCTS framework. Our algorithm demonstrates its effectiveness in guiding autonomous vehicles through real-world derived scenarios with traffic con-gestion, such as navigating roundabouts, intricate intersections, dynamic merge/splits, executing unprotected left turns, and managing cut-ins and ramps. Qualitative examples highlight the algorithm's proficiency in making diverse driving decisions, including lane changes, acceleration, and deceleration. Furthermore, the quantitative analysis underscores the efficiency of our approach, with the ability to deliver decision-making results (including longitudinal and lateral decisions) within less than one second, showcasing its potential for real-time, online planning. The success rate across various scenarios further attests to the algorithm's robustness, showcasing its ability to handle complex driving decisions and reinforcing its effectiveness and reliability.