System Research On Robot Apple Picking Path Planning Based On Deep Reinforcement Learning

In the picking and tracing control,the apple picking robot will encounter interference between the robot body and the branches,stems and leaves of the fruit,and even cause damage to the robot when it is serious.This is caused by the defects in the processing mechanism of the traditional decision control algorithm for specific scenarios.As an artificial intelligence algorithm at the forefront of solving strategic planning problems,deep reinforcement learning algorithms play an advantage in solving problems that are difficult to establish accurate mathematical models and physical control problems.Devices that use deep reinforcement learning as the main control algorithm usually have high flexibility and robustness.Therefore,it is hoped to use the control system of deep reinforcement learning to construct the path decision calculation system of the apple picking robot to solve the problem of insufficient flexibility of the picking robot.So,this paper conducts research based on the robot apple picking path planning of deep reinforcement learning.The main work of the thesis is as follows:(1)The traditional control theory system of apple picking robot is analyzed,and the three modules of perceptron,controller and decision maker of the picking robot control system are constructed through the acquisition of robot environmental information,robot control mode and picking trace strategy,which lays a foundation for the study of robot control system based on deep reinforcement learning.(2)The tracing strategy and control system construction of apple picking robot in deep reinforcement learning are investigated.First of all,it discusses what kind of computer simulation training platform to build when training the decision-making neural network of apple picking robot.In addition,the construction method of the deep neural network path decision system of the apple picking robot is studied.It also studies how to migrate the mature neural network trained by the simulation training platform into the control system of the physical robot.Finally,the experimental results of the simulation robot’s ability to obtain environmental rewards increase with the progress of training verify the feasibility of deep reinforcement learning tracking strategy.(3)In order to meet the needs of the picking robot control system to read in signals in a specific environment,an obstacle position calibration scheme suitable for deep neural network picking robot tracking is proposed.In this scheme,the robot body pose is compared with the cluster center set of fruit tree branches,and the most interfering points for the picking path of the robot are obtained,which are recorded as the obstacle points of the current path decision to participate in the subsequent calculation.At the end of the experiment,the relative success rate of 93.79% is the highest when the algorithm is used to calibrate the obstacle position of the obstacle tracked by the picking robot,which verifies the feasibility of the obstacle calibration scheme of the picking robot tracked by the deep neural network.(4)In order to avoid the failure of deep neural network path decision-making and the damage caused to the picking robot by outputting wrong actions,a safety protection scheme for deep reinforcement learning picking robot is proposed.This scheme uses multiple sets of neural networks to jointly make decisions on the state of the robot,and selects an optimal set of solutions in multiple sets of path solutions as the basis for driving the robot.And the virtual robot is built for experimental verification,and it is found that the success rate of the original path decision of the experimental picking robot can not only be increased to 93.5% by the original58.8% of the safety algorithm,but also the wrong action of the robot can be curbed,indicating that the safety algorithm can effectively improve the success rate of the path decision of the deep reinforcement learning picking robot.This paper focuses on the construction of a picking robot path control system that is more suitable for giving full play to the advantages of deep reinforcement learning in intelligent decision-making.This study has great reference value for the design of intelligent control scheme of apple picking robot for deep reinforcement learning.