Research On Robotic Arm Grabbing Method Based On Deep Reinforcement Learning

Robotic arms have always occupied an important position in industrial production,and the problem of robotic arm grasping has always been a research hotspot in academia.Traditional vision-based robotic arm grasping methods generally require a series of operations such as target recognition,target segmentation,feature matching,and coordinate reconstruction.In recent years,deep reinforcement learning has become the focus of attention in the field of artificial intelligence.Deep reinforcement learning combines the perception ability of deep learning and the decision-making ability of reinforcement learning,and has achieved breakthrough results in many high-dimensional control problems.Therefore,this article explores a method based on deep reinforcement learning to solve the continuous control problem of robotic arm grasping.The research contents of this paper are as follows:(1)This article first understands the robotic arm grasping problem from the perspective of the deep reinforcement learning model,and designs a general grasping system framework and training program based on this.The interaction data between the deep reinforcement learning agent and the robotic arm is determined:The state space of the agent is selected as the image pixel information obtained by the camera sensor of the robotic arm,and the action space of the agent is position of the robotic arm.(2)This paper explores the use of the Deep Deterministic Policy Gradient(DDPG)algorithm,which is widely used in continuous control problems in deep reinforcement learning,as an agent specific implementation on the general framework designed above.The simulation experiment is carried out in the PyBullet simulation environment,and the DDPG agent is trained to complete the grasping task of the robotic arm.Then,aiming at the problems of low sampling efficiency of DDPG algorithm and sparse rewards,a HER-DDPG algorithm combining Hindsight Experience Replay(HER)and DDPG is designed.Experiments were carried out in the same experimental environment,which improved the performance of the model and reduced the training time of the model,and finally improved the success rate of grasping of the robotic arm.(3)This paper explores the use of the on-policy algorithm of deep reinforcement learning to represent the Proximal Policy Optimization algorithm(PPO)to complete the grasping task.The PPO algorithm has the advantages of relatively few hyperparameters and stable performance.The simulation experiment was carried out in the PyBullet simulation environment,and the PPO agent was trained to complete the grasping task of the robotic arm.Then this paper designs a multi-threaded version of the PPO algorithm from the perspective of parallel computing optimization,which greatly reduces the time it takes for the algorithm to interact with the environment to collect data,and improves the efficiency of the algorithm to collect data.The optimal successful trajectories of each algorithm in the test environment are selected for comparison,and the trajectories learned by the multi-threaded version of the PPO algorithm perform best compared to the trajectories of other algorithms.