Research On Key Technologies Of Apple Picking Robot In Unstructured Environments

With the continuous development of robot technology,its application in the field of agriculture is becoming more and more extensive.For example,robots are widely used in agricultural picking.As a new type of agricultural robot,the apple picking robot mainly uses various technical means such as visual recognition,robot positioning,grasping control,etc.,and can independently complete the work of apple identification,collection and classification,so as to actually discover the fully automatic apple collection.The conventional manual collection method,apple collector personnel have the advantages of high collection efficiency,stable collection quality,and reduced labor costs.Therefore,the application of apple pickers can not only effectively solve the problem of insufficient agricultural labor power,but also has great significance in promoting the modernization and intelligence of agricultural production.This paper focuses on the application-oriented research on the key technologies of the automatic apple collection robot,focusing on the software design based on the collection robot,the target knowledge recognition technology,the extraction control technology,and the terminal highly precise control technology.The primary focus and innovative aspects of this paper are outlined below:(1)In order to solve the problem of picking apples with different growth postures,according to the actual problems such as the height of the apple,the growth environment of the fruit tree,and the force of the fruit,the software and hardware of the manipulator and gripper of the apple picking robot were tested.Design,mathematical modeling and simulation.It mainly includes the construction of the kinematics model of the picking robot arm,the construction of the camera imaging model,the hardware selection,design and software design of the picking robot.Especially in the hardware part,an apple picking robot composed of a mobile platform and a Zu7 joint coordinate manipulator composed of 6degrees of freedom is proposed.Because of its anthropomorphic structure and high flexibility,it can reach any target position in space.Meet the rigid requirements of apple picking.(2)In order to solve the problem of fine force control during the gripping process of the end effector of the robotic arm.In this paper,a soft tactile sensor is designed based on the optical principle,which consists of three parts: colored silicone,camera and light source.The sensor is not only simple in structure and small in size,but also solves the problem of interference of traditional sensor measurement data.Then,by attaching soft tactile sensors to the robot’s gripper fingertips,the robot can utilize a wealth of tactile feedback for challenging grasping tasks that require fine force control,such as stably grasping fragile objects.At the same time,by studying the interpretable image recognition mechanism of deep learning in the process of three-dimensional force decoupling and different flexible probe pattern layer production methods,it will guide the production of the fingertip flexible sensor probe of the manipulator.(3)In order to solve the problem of accurate fruit identification by picking robots in different orientations,different light intensities and at different times,this paper proposes an improved algorithm for apple recognition technology and apple recognition technology in low light conditions to meet apple recognition in complex situations.First of all,in order to speed up the extraction of features,improve the accuracy of small target detection and the overall detection speed,an improved HOG+SVM algorithm that combines the Focus+CSP module and the FPN module is proposed.The average recognition accuracy increased by13.15% under various situations such as slight occlusion and severe occlusion.Then,in order to improve the algorithm’s ability to identify and locate apples,the deep learning model based on the attention mechanism was used to obtain more abstract and obvious semantic feature information,etc.,and these features were fused into higher-level features for global analysis.information to predict.In addition,the loss function has been improved.Compared with the method of detecting targets through sliding windows,it is more efficient,has fewer false detection for the background,and is easier to achieve real-time performance.The experimental results reveal a substantial advancement in the recognition accuracy achieved by the improved algorithm,surpassing the performance of the HOG+SVM method by an impressive margin of 12.12%.Finally,in order to improve the accuracy of apple recognition under low-light conditions,an unsupervised Generative Adversarial Networks(GAN)model with a shared part structure is designed to process the input image through low-light image enhancement technology,improve the brightness,contrast and color of the image,while retaining and enhancing the detail information in the image as much as possible,and the image can be enhanced without paired training data.Because the model with shared structure is adopted,not only the size of the model can be reduced,but also the stability of training can be improved.(4)Convergence speed and robustness problems in grasping control of apple picking robot.This paper proposes an adaptive gain adjustment controller based on fuzzy logic to ensure the fast convergence and robustness of the visual servo system.Then,the apple grasping technology based on the combined visual and tactile perception of deep reinforcement learning,through the use of visual perception and tactile perception to work together in the alignment adjustment and clamping stages,established a mapping relationship between the two perception methods and improved The work efficiency of the robot in the apple picking task;finally,in the process of grasping the apple control,in order to improve the control accuracy,the contact state between the gripper and the apple is analyzed through the Gaussian mixture model(GMM)and other learning algorithms,and the classification is established.The nonlinear mapping relationship between the control variable and the state variable is established,and the Gaussian mixture regression(GMR)algorithm is used to realize the compliant control of the gripper’s pose.Experimental results show that the amplitude of trajectory generated by GMR changes less and converges faster.Finally,the sampling-based motion planning algorithms RRT and RRT* are analyzed in detail.The RRT* algorithm ensures that the robotic arm finds a better path than just a feasible path as RRT does,and usually the path is not the shortest,and the results obtained by each algorithm sampling are different.