Research On End-to-end Path Planning Method Of Autonomous Mobile Robot Based On Optical Imaging

Autonomous mobile robots have been widely used in many fields,and become more and more important.Path planning is the key for mobile robots to realize autonomous and flexible navigation,meanwhile,research on path planning methods is of great significance for ensuring the autonomy and intelligence of mobile robots.In recent years,the technology of optical imaging sensor(i.e.,camera)has been widely used in the field of robotics by virtue of its advantages in collecting environmental information and low cost.At present,traditional path planning methods usually need to construct accurate environmental maps in advance,which cannot effectively achieve path planning when facing unknown environments.These methods also have disadvantages such as cumbersome process,low efficiency,and poor robustness.Aiming at the problems of traditional methods,this study proposed an end-to-end path planning method for autonomous mobile robots based on optical imaging,with deep reinforcement learning technology.It adopts a learning method and allows the robot to continuously "trial and error" to learn an optimal movement path.By constructing a pose-based navigation dataset,and studying a pose-guided end-to-end path planning constructing method,to realized an end-toend path planning that directly learns the navigation path through captured images.The specific research content includes the following parts:First,in order to provide a basic interactive environment for the end-to-end path planning methods,a pose-based dataset is constructed.RGB images and depth images are collected by optical imaging sensors,the necessary information that can guide the robot to interact with the environment and realize path planning is then obtained according to the images: first,the pose between landmarks is directly obtained by using 3D object detection method,which realizes the association between landmarks.Then,the shortest distance between any two landmarks is obtained by Floyd-Warshall algorithm,which provides true-value reference information for the subsequent training of the path planning model.Finally,mark the basic information such as rotation by using HDF5,for guiding mobile interaction of mobile robots.Through the above methods,a flexible interactive environment is provided for the end-to-end path planning methods based on deep reinforcement learning.Secondly,to improve the efficiency and flexibility of path planning,pose-guided DRL method are used to design an end-to-end path planning model.First,the dynamic action space design is realized by transforming the pose information directly into the action of the robot,and designing a dynamic action selection mechanism.Then,the traditional reward function based on the number of moving steps are improved and a reward mechanism based on navigation costs(i.e.,rotation times and movement distance)are realized,which further reduce the cost of robot movement and improve navigation efficiency.Finally,a pose-guided end-to-end path planning model is proposed by improving the Deep Siamese Actor-Critic(Deep Siamese ActorCritic)DRL network model with dynamic action space,navigation cost-oriented reward mechanism.The optimal decision-making strategy for action of mobile robot is finally generated by using the proposed network model,which has been trained to interact with the environment using dataset,to guide the mobile robot to choose the direction of movement.Through the above research,the efficient and flexible end-to-end path planning of the mobile robot is realized.Finally,to verify the performance of the proposed optical imaging based end-to-end path planning method,a series of comparative experiments are carried out.First,multiple common scenarios such as laboratories,living rooms,and apartments are selected to construct a mobile robot navigation interactive environment using the proposed pose-guided dataset construction method,to provide a test environment for following experiments.Then,the efficiency and flexibility of navigation as well as the comprehensive performance of the end-to-end path planning method are verified.The experimental results show that the proposed method has higher efficiency and better flexibility compared with the current advanced end-to-end path planning methods,proving the great performance of the end-to-end path planning method proposed in this study.