Scholarlink

<p>Given the prevalent limitations of existing deep reinforcement learning-based fusion methods, including heavy reliance on environmental priors, grid maps, complex cross-modal alignment, and strong sensor coupling, this paper proposes a mapless deep reinforcement learning (DRL) framework. The framework fuses visual and 2D LiDAR data, integrates a self-attention soft actor-critic (SAC) architecture and a customized reward function, and achieves robust autonomous navigation and exploration without requiring obstacle priors. First, a comprehensive vision-augmented LiDAR fusion technique is proposed that enhances obstacle perception using RGB images, enabling efficient processing of high-dimensional sensor inputs. Second, a self-attention Soft Actor-Critic (SAC) framework is designed to capture long-range spatial dependencies in sensor data through a feature extractor, improving policy robustness in cluttered scenes. In addition, this paper customizes a dedicated reward function, which dynamically balances exploration efficiency, obstacle avoidance performance, and goal-directed navigation capability in a mapless manner. Finally, extensive Gazebo simulation and real robot experiments are conducted to demonstrate the robustness and adaptability of the proposed method in tackling the complexities of autonomous navigation in complex and unpredictable environments.</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

Self-attention SAC with vision-augmented LiDAR fusion for mapless robot navigation in dynamic environments

  • 期刊论文

  • 发表日期：

    2026年7月25日

查看全文
  • Heng Deng,
  • Boyu Cui,
  • Tong Jia,
  • Jingyuan Zhan,
  • Wentao Jiang,
  • Liangji Shen,
  • Liguo Zhang

摘要

Given the prevalent limitations of existing deep reinforcement learning-based fusion methods, including heavy reliance on environmental priors, grid maps, complex cross-modal alignment, and strong sensor coupling, this paper proposes a mapless deep reinforcement learning (DRL) framework. The framework fuses visual and 2D LiDAR data, integrates a self-attention soft actor-critic (SAC) architecture and a customized reward function, and achieves robust autonomous navigation and exploration without requiring obstacle priors. First, a comprehensive vision-augmented LiDAR fusion technique is proposed that enhances obstacle perception using RGB images, enabling efficient processing of high-dimensional sensor inputs. Second, a self-attention Soft Actor-Critic (SAC) framework is designed to capture long-range spatial dependencies in sensor data through a feature extractor, improving policy robustness in cluttered scenes. In addition, this paper customizes a dedicated reward function, which dynamically balances exploration efficiency, obstacle avoidance performance, and goal-directed navigation capability in a mapless manner. Finally, extensive Gazebo simulation and real robot experiments are conducted to demonstrate the robustness and adaptability of the proposed method in tackling the complexities of autonomous navigation in complex and unpredictable environments.