This paper presents a novel robotic navigation system integrating a visual simultaneous localization and mapping (V-SLAM) based global planner with a deep reinforcement learning (DRL) based local planner. On one hand, map of many modern popular V-SLAM systems is inhomogeneous point cloud, which contains many outliers and is too sparse for reliable global path planning. To address this problem, we propose a novel approach to generate a topological map with both trajectories and map points of V-SLAM. On the other hand, current state-of-the-art (SOTA) DRL-based local planners have shown great efficiency in obstacle avoidance. However, the SOTA DRL-based local planners are sometimes trapped by large obstacles and would fall into some local minimum during training. To address the problems, we propose a sub-target module and a mirror experience replay approach. Test results demonstrate that, our topological map generation is robust against outliers and sparsity of map points of V-SLAM, while our local planner achieves 9.61% success rate of obstacle avoidance higher than the SOTA DRL-based approach. Tests in real environment demonstrate the feasibility of our navigation system.

本文提出了一种新型机器人导航系统，它集成了基于视觉同步定位和映射 (V-SLAM) 的全局规划器和基于深度强化学习 (DRL) 的局部规划器。一方面，许多现代流行的 V-SLAM 系统的地图是不均匀的点云，其中包含许多异常值，并且对于可靠的全局路径规划来说过于稀疏。为了解决这个问题，我们提出了一种新方法来生成具有 V-SLAM 的轨迹和地图点的拓扑图。另一方面，当前最先进的（SOTA）基于 DRL 的本地规划器在避障方面表现出很高的效率。然而，基于 SOTA DRL 的局部规划器有时会被大障碍困住，并且在训练过程中会陷入一些局部最小值。为了解决问题，我们提出了一个子目标模块和一个镜像体验重放方法。测试结果表明，我们的拓扑图生成对 V-SLAM 的异常值和地图点的稀疏性具有鲁棒性，而我们的本地规划器实现了 9.61% 的避障成功率，高于基于 SOTA DRL 的方法。在真实环境中的测试证明了我们导航系统的可行性。