Research On Intelligent Control Method Of Robotic Arm Based On Improved DDPG Algorithm

With the rapid development of robotics,the use of robotic arms is becoming more and more widespread.There are limitations in traditional robotic arm control methods,which require precise mathematical models based on the task,rely heavily on manual operations,are cumbersome and lack self-adaptability,and are less effective if the task or environment of the robot arm changes.In recent years,intelligent control algorithms have been a hot topic of research in robot control,and one important direction in the field of artificial intelligence is deep reinforcement learning,which has been successfully applied to many aspects of robotic arm control,giving the robot arm the ability to learn and interact with the environment on its own,making up for the shortcomings of traditional control methods.However,the dynamics of the working environment in which the robotic arm operates and the complexity of the model in which it operates leave many problems with the application of intelligent algorithms to robotic arm control.To this end,this paper focuses on the target arrival and target grasping tasks in robotic arm control based on a deep reinforcement learning approach,and proposes improvements to the intelligent control algorithm in order to improve its control performance in the field of robotic arms.This paper uses the DDPG algorithm to implement a control study of a robotic arm with an unknown environment.To address the problems of sparse rewards and low data sampling utilization of the DDPG algorithm,an improved DDPG algorithm is proposed to improve these two aspects respectively.A hybrid reward function with sparse rewards,distance rewards,directional rewards and area rewards superimposed on each other is designed to improve the training effect of the algorithm.On the problem of sampling data utilization efficiency of the experience replay mechanism,a new experience pool is added to store the sample data in descending order of reward value,and the data ranges of priority sampling and uniform sampling are proportionally divided to collect samples,so that more high-quality data are used for the update of network weights and optimization parameters of the algorithm.The robotic arm is studied in depth on a two-dimensional simulation platform with an autonomous control end to achieve the task of target object arrival.The improved DDPG algorithm before and after training in a simulation environment,and the comparison results of the experiments show that improvements in both the reward function and the experience playback mechanism can improve the stability and convergence speed of the robotic arm in completing the task,and the integrated improved DDPG algorithm shows better results in the control performance of the robotic arm.The robotic arm target arrival and grasping control tasks are investigated on a three-dimensional seven-degree-of-freedom robotic arm simulation platform in which the algorithm improvements are migrated.In the robotic arm target arrival task,it is further verified that both improvements lead to improved learning efficiency and success rate of the robotic arm.In the robotic arm grasp control task,the algorithm uses the observed environment image as input and the continuous action space at the end of the robotic arm in the spatial coordinate system as the output of the decision based on the ability of the convolutional neural network to extract features from the image.By designing a hybrid reward function,the robotic arm is able to control the grasp of the target object more accurately.The experience replay mechanism is improved and different capacity experience pools are set.Experimental results show that the capacity size of different experience replay pools has an impact on the grasping effect of the robotic arm,and the improved algorithm has a higher grasping success rate and shows better control performance.The comprehensive improved DDPG algorithm can achieve88.96% success rate in the grasping task.The improved solution facilitates the robotic arm to learn control strategies to complete specific tasks autonomously and improves the robotic arm’s environmental self-adaptation capability.