DRL-Based Dynamic Path Planning for Multi-sensor Mobile Robot

In the open-air logistics park, mobile robots need to navigate autonomously and avoid dynamic obstacles in real time to move goods between warehouses and loading places. To solve this problem, this paper presents a path planning method for mobile robot based on deep reinforcement learning (DRL). The method uses two neutral network structures - long short-term memory network (LSTM) and convolutional neural network (CNN). LSTM is used to process robot’s own state information and LiDAR point cloud information. CNN is used to extract features of image information to identify dynamic obstacles and target point in the environment, semantic segmentation is also used to remove two unrelated factors of sky and ground with masks. Combining these data sources, the dueling double deep Q-Network with prioritized experience replay (D3QN PER) algorithm directly maps sensor information to action instructions. Simulation results demonstrate that the proposed method endows mobile robots with enhanced adaptability and environmental understanding capabilities. Moreover, this method is effective when the dynamic scene changes, outperforms APF algorithm.