Scholarlink

<p>This study explores cooperative control among multiple autonomous vehicles of varying sizes in a shared environment using deep reinforcement learning (DRL). Although our previous study demonstrated the feasibility of cooperative control in such mixed-size environments, it did not incorporate vehicle size information, meaning that vehicles were unaware of their own sizes or those of others. In this study, we propose a DRL method that explicitly considers vehicle size to enable more effective cooperative control. We integrate two approaches to utilizing size information: using only the vehicle’s own size, and using both the vehicle’s own size and neighboring vehicles’ sizes. Three encoding methods for representing size are also investigated. Simulation experiments are conducted using standard passenger cars, medium-duty trucks, and compact vehicles. Models are trained in two environments with different levels of vehicle type diversity and evaluated on unseen course layouts. Performance is assessed by collision rates and average lap times. The results show that incorporating vehicle size information improves generalization, especially when using only the vehicle’s own size and a relative encoding. This configuration achieves more stable cooperative behavior in environments with previously unseen vehicle types.</p>

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

Investigating vehicle size input strategies in deep reinforcement learning for cooperative autonomous driving

  • 期刊论文

  • 发表日期：

    2025年12月29日

查看全文
  • Akito Takenaka,
  • Tomohiro Harada,
  • Yukiya Miura,
  • Kiyohiko Hattori,
  • Johei Matuoka

摘要

This study explores cooperative control among multiple autonomous vehicles of varying sizes in a shared environment using deep reinforcement learning (DRL). Although our previous study demonstrated the feasibility of cooperative control in such mixed-size environments, it did not incorporate vehicle size information, meaning that vehicles were unaware of their own sizes or those of others. In this study, we propose a DRL method that explicitly considers vehicle size to enable more effective cooperative control. We integrate two approaches to utilizing size information: using only the vehicle’s own size, and using both the vehicle’s own size and neighboring vehicles’ sizes. Three encoding methods for representing size are also investigated. Simulation experiments are conducted using standard passenger cars, medium-duty trucks, and compact vehicles. Models are trained in two environments with different levels of vehicle type diversity and evaluated on unseen course layouts. Performance is assessed by collision rates and average lap times. The results show that incorporating vehicle size information improves generalization, especially when using only the vehicle’s own size and a relative encoding. This configuration achieves more stable cooperative behavior in environments with previously unseen vehicle types.