Research On Autonomous Learning And Training Of Robot Picking Action

With the continuous development of artificial intelligence and robot technology,intelligent mechanical equipment such as robots and mechanical arms are gradually applied to industrial and agricultural production and manufacturing.Especially in terms of agricultural picking,taking cherry tomato planting industry as an example,cherry tomato picking has a large workload,high cost,and labor shortage.However,most of the current cherry tomato picking robots are in the laboratory stage and do not retain the fruit pedicle.The picking method is difficult to meet the needs of the domestic market.In order to solve the above problems,this paper takes the intelligent picking of cherry tomatoes as an example,according to the growth characteristics of cherry tomatoes and the skills of manual picking,a set of plans for imitating manual picking of cherry tomatoes was formulated,that is,the end effector of the mechanical arm can keep basically the same direction as the fruit pedicle of the cherry tomato when picking,so that the subsequent end effector can pick the cherry tomato with the fruit pedicle in a way of folding the fruit pedicle.Here we focus on a specific picking action generation system,combining robotic arm control with machine vision,and using deep reinforcement learning technology to allow the robotic arm to learn autonomously during training and automatically generate the specific robotic arm picking action according to the direction of the cherry tomato pedicle.(1)Obtain the pose information of the target fruit through machine vision.Design and implement the ripe tomato target recognition algorithm;according to the idea of the straight line fitting algorithm,relevant improvements are made,and the contour characteristics of the cherry tomatoes are fully highlighted through the image thinning operation,and the direction data of the cherry tomatoes are obtained by using the improved direction fitting algorithm;using the size relationship of the object image to realize the horizontal forward distance measurement method of the monocular camera,and the accuracy,stability and feasibility of the algorithm have been verified through multiple tests.(2)Research on robotic arm motion training based on deep reinforcement learning.Model and simulate the domestic Gauss six-axis manipulator,and build a simulation environment.Analyze the advantages and disadvantages of related reinforcement learning algorithms,and choose the deep deterministic policy gradient algorithm as the basic algorithm.The algorithm is adaptively improved in terms of increasing the state dimension and designing a targeted hierarchical reward function,which reduces the difficulty of reward acquisition and improves the convergence speed of the algorithm.The simulation test was carried out in Gazebo,and the result reached the expected goal.(3)Build and test the picking action generation system.Build the picking system from the hardware aspect,and deploy the ROS-based communication mechanism on the software aspect accordingly;the basic algorithm modules such as the target fruit recognition module,the fruit pose analysis module,the distance measurement module,and the action generation module are tested and analyzed respectively,and the feasibility conclusion of the algorithm is drawn;the overall system is tested to generate specific picking actions according to the target fruit pose and the experimental data in the process is recorded.The data shows that the errors between the algorithms are superimposed,and the success rate of specific actions is about 84%.