Scholarlink

This paper presents a simulation-based study on securing an autonomous driving system using deep reinforcement learning (DRL) techniques, particularly Deep Q-Network (DQN) and Double Deep Q-Network (DDQN). The goal is to design an intelligent and safe driving agent capable of maintaining optimal speed, avoiding collisions, and ensuring stability in dynamic environments. The CARLA simulator is used as the testbed, with a convolutional neural network (CNN) employed for state representation and policy learning. Safety is integrated at multiple levels: reward shaping penalizes risky behavior, lane detection is enhanced through the Canny edge detection algorithm, and vehicle parking is supported by geometric calculations to minimize collisions. Experimental results show that DQN achieves a smoothed accuracy of 0.80 with a peak of 0.81, while DDQN demonstrates better consistency with a smoothed accuracy of 0.92 and the same peak performance. These findings confirm the potential of DRL approaches in developing secure and efficient autonomous driving systems suitable for real-world deployment.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Secure an Autonomous Driving System Using Deep Reinforcement Learning: A Simulation-Based Study in CARLA

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mohamed Khayati,
  • Mohamed Ouaskou,
  • Mohamed Baslam

摘要

This paper presents a simulation-based study on securing an autonomous driving system using deep reinforcement learning (DRL) techniques, particularly Deep Q-Network (DQN) and Double Deep Q-Network (DDQN). The goal is to design an intelligent and safe driving agent capable of maintaining optimal speed, avoiding collisions, and ensuring stability in dynamic environments. The CARLA simulator is used as the testbed, with a convolutional neural network (CNN) employed for state representation and policy learning. Safety is integrated at multiple levels: reward shaping penalizes risky behavior, lane detection is enhanced through the Canny edge detection algorithm, and vehicle parking is supported by geometric calculations to minimize collisions. Experimental results show that DQN achieves a smoothed accuracy of 0.80 with a peak of 0.81, while DDQN demonstrates better consistency with a smoothed accuracy of 0.92 and the same peak performance. These findings confirm the potential of DRL approaches in developing secure and efficient autonomous driving systems suitable for real-world deployment.