Scholarlink

<p>Deep reinforcement learning (DRL) has emerged as a powerful tool for autonomous robot navigation, enabling robots to adapt to dynamic environments through interactive learning. Despite extensive research on DRL model architectures and policy optimization techniques, the influence of critical hyperparameters, particularly the LR and optimizer selection, on navigation performance remains underexplored. Addressing this gap is essential for advancing reliable and efficient DRL-based navigation systems. This study systematically evaluates how learning rates and optimizer choices influence the training stability, learning efficiency, and navigation success of DRL-based mobile robots. Three algorithms were investigated: DQN, DDPG, and TD3. A ROS2-based DRL navigation stack was developed to test LR and optimizers such as Adam, RMSprop, and SGD across the selected algorithms. Performance metrics included navigation success rate and learning efficiency in dynamic environments. Our experiments demonstrate that both the optimizer and LR substantially impact the navigation performance of DRL models. Careful tuning of these hyperparameters significantly enhances training stability and the robot’s navigation success rate. In particular, lower LR combined with adaptive optimizers such as RAdam led to more stable training and higher success rates, especially with TD3 and DDPG. These configurations also improved learning efficiency and reduced training instability. This study underscores the critical role of hyperparameters in DRL-based navigation systems. The findings provide actionable guidelines for selecting optimizers and LR to enhance decision-making, adaptability, and robustness in real-world robotic applications. These insights help bridge a significant gap in DRL, offering a pathway to deploy more reliable autonomous navigation solutions.</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

Enhancing Robot Navigation in Crowded Environments with Deep Reinforcement Learning

  • 期刊论文

  • 发表日期：

    2025年8月16日

查看全文
  • Fatma Zohra Ouamane,
  • Foudil Cherif,
  • Meftah Zouai

摘要

Deep reinforcement learning (DRL) has emerged as a powerful tool for autonomous robot navigation, enabling robots to adapt to dynamic environments through interactive learning. Despite extensive research on DRL model architectures and policy optimization techniques, the influence of critical hyperparameters, particularly the LR and optimizer selection, on navigation performance remains underexplored. Addressing this gap is essential for advancing reliable and efficient DRL-based navigation systems. This study systematically evaluates how learning rates and optimizer choices influence the training stability, learning efficiency, and navigation success of DRL-based mobile robots. Three algorithms were investigated: DQN, DDPG, and TD3. A ROS2-based DRL navigation stack was developed to test LR and optimizers such as Adam, RMSprop, and SGD across the selected algorithms. Performance metrics included navigation success rate and learning efficiency in dynamic environments. Our experiments demonstrate that both the optimizer and LR substantially impact the navigation performance of DRL models. Careful tuning of these hyperparameters significantly enhances training stability and the robot’s navigation success rate. In particular, lower LR combined with adaptive optimizers such as RAdam led to more stable training and higher success rates, especially with TD3 and DDPG. These configurations also improved learning efficiency and reduced training instability. This study underscores the critical role of hyperparameters in DRL-based navigation systems. The findings provide actionable guidelines for selecting optimizers and LR to enhance decision-making, adaptability, and robustness in real-world robotic applications. These insights help bridge a significant gap in DRL, offering a pathway to deploy more reliable autonomous navigation solutions.