Scholarlink

For the problem that existing deep reinforcement learning (DRL) methods are ineffective in navigation tasks of complex environments containing structures such as long corridors and corners, a hierarchical deep reinforcement learning (HDRL) framework is proposed. The HDRL framework includes two low-level models which generate goal-driven actions and collision avoidance actions separately, and a high-level model, which learns a reasonable strategy to choose between the two actions, thus avoiding the reliance on the priori knowledge in the design of the reward functions. In addition, a subgoal is selected to reduce the problem of sparse reward and improve generalization ability for the DRL model. The HDRL method outperforms existing DRL methods at all indexes in the comparative experiments.

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

Hierarchical Deep Reinforcement Learning for Path Planning with Collision Avoidance of the Mobile Robot in Complex Dynamic Environments

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yuanyuan Gao,
  • Qingling Wang

摘要

For the problem that existing deep reinforcement learning (DRL) methods are ineffective in navigation tasks of complex environments containing structures such as long corridors and corners, a hierarchical deep reinforcement learning (HDRL) framework is proposed. The HDRL framework includes two low-level models which generate goal-driven actions and collision avoidance actions separately, and a high-level model, which learns a reasonable strategy to choose between the two actions, thus avoiding the reliance on the priori knowledge in the design of the reward functions. In addition, a subgoal is selected to reduce the problem of sparse reward and improve generalization ability for the DRL model. The HDRL method outperforms existing DRL methods at all indexes in the comparative experiments.