Large-scale Autonomous Navigation and Path Planning of Lunar Rover via Deep Reinforcement Learning

Future lunar surface exploration needs to be carried out in areas with more complex terrain and richer scientific research value without prior information. Therefore, the patrol device needs to be able to autonomously detect terrain and obstacle avoidance when planning exploration tasks. However, Deep reinforcement learning has been successfully applied in various game-like environments, and it is a challenging task to apply deep reinforcement learning to the navigation and obstacle avoidance detection of the rover on lunar surface. We propose an autonomous decision-making method for lunar vehicle patrols based on deep reinforcement learning. The algorithm first uses lidar for simultaneous positioning and graphing. For the constructed static three-dimensional point cloud map of the lunar surface, it is described as a two-dimensional grid using octree-map projection The grid map is processed by the long-short time memory (LSTM) to process the two-dimensional grid obstacle information. Using dueling deep Q network (dueling DQN) with multi-step learning to train the planetary vehicle automatic path planning, and finally in the ROS static obstacle environment simulation verification experiment, the simulation results verify that the proposed method is effective and adaptable for the lunar rover in different terrain environments.