The Research Of Mobile Robot Navigation Based On Safety Priority Guidance

With the continuous development of intelligent robot technology,the development degree of robot technology has gradually become one of the standards to judge the strength of a country’s advanced scientific and technological level.A very important link in the research of intelligent robot technology is the navigation technology of mobile robot,which concerns whether the robot can achieve the basic function of navigation to the destination according to the requirements.In this study,we propose a safety-oriented mobile robot navigation method based on deep reinforcement learning,visual SLAM,semantic segmentation network,3D target detection and so on.This method is aiming at the problem that navigation security is seldom paid attention to in traditional mobile robot navigation research.From 3D scene understanding and construction to 2d grid map containing safety information,safety priority navigation is implemented by natural language control.This study mainly includes the following research work:(1)Safe-PPO algorithm based safe-priority path planning methodA safe-PPO algorithm for mobile robot path planning is proposed.Based on the PPO algorithm in deep reinforcement learning,the PPO algorithm is optimized by referring to the idea of evolutionary strategy,and two safety-related parameters of hazard coefficient and movement coefficient are introduced to design the safety reward function.This study constructed a two-dimensional grid map environment and did various comparative experiments in this environment.This study also did three realworld experiments to verify the proposed algorithm.The experimental results show that the algorithm proposed in this study is reasonable and feasible,and the robot can gradually learn the safety-oriented path planning skills,and choose the safer path instead of the closer path to the target point in the experimental environment.(2)Semantic octmap SLAM algorithm based on improved HRNetIn this study,the visual SLAM system was combined with semantic segmentation network.The camera pose estimation provided by ORB-SLAM2 and pixel-level classification provided by the improved HRNet semantic segmentation network were used to construct semantic point cloud,and the maximum probability fusion algorithm and voxel filtering were used to construct 3D semantic octree map.In the improvement of HRNet network,combined with the parallel hierarchical structure of HRNet,multilayer serial dilated convolution structure and multi-layer DUpsampling structure are designed in this study.When constructing 3D semantic map,the maximum probability fusion algorithm is used to process semantic information in semantic point cloud,and then the 3D semantic octmap is formed by voxel filtering.The performance test of semantic segmentation network proves that the improved HRNet network has better accuracy.The experiment of map construction shows that the proposed map construction method can effectively construct 3d environment map containing semantic information and save the map to various resolutions.(3)Research on safety information fusion grid map construction and natural language safety priority navigationOn the basis of the previous two studies,a grid map construction method integrating safety information and a safety priority navigation method based on natural language input in this map are proposed.For 2D map construction,3D semantic octmap is projected onto 2D grid map by multi-layer fusion projection algorithm,and semantic information is retained.For the navigation algorithm with natural language as input,firstly,Flash Text algorithm is used to extract keywords and update navigation targets for input message.Then safe-PPO algorithm is used to carry out global path planning,and the cost map and local path planning are updated in progress.Finally,navigation to the vicinity of the target object.In the experimental verification,the warehouse scene in the TUM-RGBD dataset was used to construct the map,and the natural language command navigation was carried out in the constructed map.In the experiment,the robot can extract keywords of natural language navigation instructions and find safety priority roads to navigate around objects in the map.In this study,a safety priority navigation method for mobile robots is proposed.Safe-PPO algorithm is used for safety priority path planning.In the problem of map construction,3D environment map is constructed with the data collected by depth camera as input and semantic segmentation is carried out,then 3D semantic octree map is built.Next,the 3D octmap is projected to 2D grid map.Finally,the navigation keywords are extracted by Flash Text algorithm,and safe-PPO algorithm is used to carry out safe-priority navigation in 2D map containing semantic information.