Scholarlink

In modern robotic systems, path planning is a critical technology. This study enhances the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm by incorporating parametric noise and a dynamic reward mechanism to improve navigational performance in complex environments. Structured noise enhances the policy search process by increasing the randomness of exploration, while the dynamic reward mechanism adjusts the reward function based on environmental feedback and task progress to more accurately guide the learning process. Furthermore, this paper introduces an advanced obstacle recognition and tracking approach utilizing the Hungarian algorithm and Intersection Over Union (IOU) scores. A novel collision risk assessment model, which considers both collision time and distance, is also developed to provide a more comprehensive evaluation of potential hazards. Through a series of simulation experiments, this paper demonstrates the proposed method’s advantages in improving obstacle avoidance efficiency and reducing training duration compared to the traditional TD3 algorithm, exhibiting strong generalization capabilities. The results indicate that the TD3 algorithm, combined with parametric noise and dynamic rewards, significantly enhances path planning outcomes, showing great potential for application. The research not only expands the application scenarios of the TD3 algorithm but also offers a new approach to robot path planning in complex dynamic environments.

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

Effective Path Planning in Dynamic Environments: A Deep Reinforcement Learning Approach with Parametric Noise

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Xinzhan Hong,
  • Yexing Yang,
  • Qieshi Zhang,
  • Meng Zhao,
  • Ziliang Ren,
  • Zhanhao Liang,
  • Jing Ji

摘要

In modern robotic systems, path planning is a critical technology. This study enhances the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm by incorporating parametric noise and a dynamic reward mechanism to improve navigational performance in complex environments. Structured noise enhances the policy search process by increasing the randomness of exploration, while the dynamic reward mechanism adjusts the reward function based on environmental feedback and task progress to more accurately guide the learning process. Furthermore, this paper introduces an advanced obstacle recognition and tracking approach utilizing the Hungarian algorithm and Intersection Over Union (IOU) scores. A novel collision risk assessment model, which considers both collision time and distance, is also developed to provide a more comprehensive evaluation of potential hazards. Through a series of simulation experiments, this paper demonstrates the proposed method’s advantages in improving obstacle avoidance efficiency and reducing training duration compared to the traditional TD3 algorithm, exhibiting strong generalization capabilities. The results indicate that the TD3 algorithm, combined with parametric noise and dynamic rewards, significantly enhances path planning outcomes, showing great potential for application. The research not only expands the application scenarios of the TD3 algorithm but also offers a new approach to robot path planning in complex dynamic environments.