Research On Obstacle Avoidance Motion Planning Of Citrus Harvesting Robot In Unstructured Environment

China is a big country of citrus production.In the process of fruit cultivation and harvesting,fruit harvesting takes up a lot of manpower.As a seasonal fruit,citrus harvesting requires intensive manual labor in a short time.It needs to find a large number of labor force in a short time.In China with the increasing aging,the shortage of labor force is increasingly restricting the citrus industry.With the development of orchard management,more and more orchard managers are seeking to fill the labor gap through mechanized,automated and intelligent citrus harvesting equipment.With the rapid development of artificial intelligence disciplines such as in-depth learning,motion planning theory and computer computing speed,harvesting robots are expected to replace human labor for citrus harvesting.In the process of harvesting,whether the manipulator can avoid branches and reach the position of Citrus decides whether the citrus harvesting robot can successfully complete the harvesting operation,so the research on the motion planning of the manipulator is of great significance.This paper focuses on the motion planning of the manipulator.Three commonly used motion planning algorithms of the manipulator,RRT,PRM and RRT-connect,are compared experimentally with the planning success rate under the planning time and specific time threshold.Finally,RRT-connect is selected as the basic motion planning algorithm of the citrus harvesting robot.When the manipulator needs to work inside the crown,many branches form a closed polygon.In this case,the planning time of RRT-connect is greatly increased compared with that of avoiding only one branch.In view of this situation,through the analysis of the configuration space in this case,the RRT-connect algorithm is improved,and a two-stage RRT-connect algorithm based on informed knowledge guidance points of configuration space(IGPRRT-connect)is proposed.In view of the fact that IGPRRT-connect only performs well when obstacles are closed or approximate closed,a parallel algorithm of RRT-connect and IGPRRT-connect is proposed,which solves the problem of poor adaptability of the two algorithms in different environments.As the precondition of obstacle avoidance motion planning of manipulator,this paper also studies the acquisition and modeling of branch obstacle information.In this paper,we use deep learning convolution neural network Mask R-CNN to identify branches and trunks.The identified branches and trunks are framed with the minimum external moment.In this paper,we propose a branch representation method of "two points + radius" : two points are the midpoint at both ends of branches and half of trunks.Diameter is branch radius.Then the "two points + radius" of the identified branch segment is obtained by Kinect V2 camera,which provides obstacle information for the obstacle avoidance motion planning of the manipulator.The simulation and control system of Citrus harvesting robot was built,and the obstacle avoidance experiment was carried out.The experiment shows that the citrus harvesting robot has the preliminary motion planning ability.