Home // International Journal On Advances in Intelligent Systems, volume 18, numbers 1 and 2, 2025 // View article

Maneuver-Based Decision Making in Autonomous Driving via Reinforcement Learning and Simulation-to-Real Transfer

Authors:
Xin Xing
Morteza Molaei
Sebastian Ohl

Keywords: Autonomous Driving; Maneuver-based Decision-making; Reinforcement Learning; Simulation-to-real transfer

Abstract:
Autonomous driving requires decision-making systems that can safely handle complex, dynamic traffic scenarios. This paper presents a maneuver-based decision-making framework for autonomous vehicles using Deep Reinforcement Learning (DRL). The approach focuses on high-level driving maneuvers, specifically Adaptive Cruise Control (ACC) for maintaining safe distances and speeds, and Automatic Emergency Braking (AEB) for collision avoidance. Policies are trained in a high-fidelity simulation environment using state-of-the-art Reinforcement Learning(RL) algorithms—Proximal Policy Optimization (PPO) and Deep Deterministic Policy Gradient (DDPG)—to learn optimal driving strategies. The learned policies are then transferred from simulation to a physical autonomous vehicle platform to evaluate real-world performance. Experiments in simulation demonstrate that both PPO and DDPG achieve efficient and safe driving behavior: DDPG converges faster and produces smoother control actions, while PPO learns more conservative policies that prioritize safety in unpredictable conditions. Real-world validation experiments corroborate effective simulation-to-real policy transfer, with PPO maintaining robust safety margins and DDPG executing more aggressive yet efficient maneuvers. In summary, the study achieves a reliable RL-based decision-making system for autonomous driving and provides a comparative analysis of policy optimization methods. Key contributions include a maneuver-based RL framework, demonstration of effective sim-to-real policy transfer, and insights into the trade-off between safety and efficiency for different RL algorithms in autonomous driving.

Pages: 46 to 56

Copyright: Copyright (c) to authors, 2025. Used with permission.

Publication date: June 30, 2025

Published in: journal

ISSN: 1942-2679