Abstract: At present,most mine inspection robots use LiDAR as the detection method of mine environment,which is not obvious for some small target objects and objects with small albedo,and is easy to cause false detection or missing detection,resulting in mine safety accidents. In order to improve the recognition accuracy of the mine inspection robot,the path planning method based on reinforcement learning combined with large neighborhood search was introduced into the path planning work of the mine inspection robot to improve the scene perception ability of the mine inspection robot. Firstly,a sequential path planning model based on LSTM is proposed,which can extract image features from the RGB camera of the robot and carry out scene perception through deep learning. Secondly,the information collected by the LiDAR equipment is processed,and the large neighborhood search algorithm is used to find multiple optimal paths in the space for the navigation of the subsequent scene. Finally, deep reinforcement learning and large neighborhood search methods are used to achieve accurate navigation of mine inspection robots,and the best inspection path is selected. In order to verify the performance of the proposed algorithm,scene construction,navigation simulation,model training and testing are carried out in 2D and 3D space. The results show that this method has better capability of path planning in simulated environment and real scene.

Key words: deep learning,large neighborhood search,time series,robotics,path planning