Research On Path Planning Method Of Kiwifruit Picking Robot Based On Deep Reinforcement Learning

With the continuous increase of kiwifruit industry and production,manual picking has gradually failed to meet market demand.The development of kiwifruit picking robots can improve picking efficiency and reduce labor costs,which is of great significance for promoting agricultural modernization and intelligent development.In order to improve the navigation and operation efficiency of the kiwifruit picking robot,this paper studies the coverage path planning and global path planning algorithm of the mobile chassis of the kiwifruit picking robot.The main work and innovations are as follows:(1)Kiwifruit projection and picking area division: Firstly,the environmental point cloud information of the kiwifruit orchard is collected by lidar,and a two-dimensional grid map of the orchard is constructed.Then a coordinate projection method of kiwifruit is proposed,and the coordinates of kiwifruit are projected into the grid map to obtain the distribution of kiwifruit in the orchard environment.Finally,a kiwifruit fruit picking area division algorithm is proposed,which combines the effective picking area of the picking robot to divide the fruit into different picking areas,so that the picking robot can carry out picking operations in targeted areas.The traditional grid-based coverage path planning method is transformed into the traveling salesman(TSP)problem for solving the traversal order of each picking area.(2)Covering path planning based on deep reinforcement learning: In order to solve the optimal traversal order of each picking area by the kiwifruit picking robot,an improved deep reinforcement learning algorithm is proposed,which is defined as the Re-DQN algorithm.Firstly,the path quality evaluation function is set,and the length of the covered path is used as the evaluation index,and the coverage path planned for each round is comprehensively scored,and the score is used as the standard for subsequent reward distribution.Then,in order to enhance the global relevance of the model,an experience repository is set to update the action value of each round as a whole,so as to speed up the learning efficiency of the model in the initial stage.Finally,set up the experience backtracking mechanism,select the high-scoring path for experience backtracking,and increase the positive guiding effect of the high-scoring path on the model.The model training results show that the Re-DQN algorithm can converge at a better solution faster than the DQN algorithm.The kiwifruit orchard navigation experiment results show that the coverage path planning method proposed in this study can effectively shorten the coverage path length of the kiwifruit picking robot and improve navigation efficiency.(3)Global path planning based on the improved RRT algorithm: The travel sequence of the kiwi fruit picking robot to each picking area is determined through coverage path planning.In order to further improve the global path planning efficiency of the kiwifruit picking robot among the picking areas,the problems of low node utilization,high path complexity,and slow convergence speed of the RRT algorithm were addressed.An improved method based on sampling state to guide random tree expansion in real time is proposed,which is defined as Straight-RRT algorithm.First,set the evaluation index and threshold to divide the sampling state of the random tree,and the sampling state determines the selection method of the sampling node.Then,improve the selection mechanism of the nearest node and set a dynamic threshold,so that the random tree can avoid obstacles faster and improve the adaptability of the algorithm to different environments.Finally,the redundant points of the path are removed and the Bezier curve is used to fit the path to obtain a smooth path with low complexity and more suitable for robot walking.The path planning experiment was carried out based on the scaffolding kiwifruit orchard environment.The experimental results show that the Straight-RRT algorithm has better adaptability and planning efficiency than the RRT algorithm,target gravity-RRT algorithm,and RRT-Connect algorithm in the kiwifruit orchard environment.