Grasping Theory And Simulation Technology Of Space Manipulator Based On Deep Reinforcement Learning

With the further expansion of space activities,conventional space manipulator control methods designed based on mission requirements are increasingly difficult to meet the needs of diverse and complex task scenarios.Meanwhile,artificial intelligence algorithms represented by deep reinforcement learning have also been introduced into the field of robot control and carried out extensive research and exploration.The application of deep reinforcement learning to space manipulator control will be a developing trend in the future.In this paper,grasping theory and simulation technology of space manipulator based on deep reinforcement learning are taken as the research topic,and the overall scheme design,experimental platform construction,algorithm implementation,scene application and other aspects are studied.The main contents are as followsFirstly,the overall scheme framework of space manipulator simulation research based on deep reinforcement learning was proposed,and the mainstream deep reinforcement learning framework,algorithm and simulation software were applied to space manipulator grasping simulation task.Secondly,a space manipulator task simulation experiment platform was developed.Based on the characteristics of MuJoCo simulation platform,two kinds of manipulator simulation scenarios of fixed base and floating base were developed.Meanwhile,aiming at the deficiency that the built-in controller of MuJoCo platform could not simulate the position tracking control of real manipulator,A joint control scheme of the manipulator was proposed based on trajectory planner and trajectory tracker,and the control scheme of the joint and claw of the manipulator was determined.Third,this paper puts forward the space manipulator trajectory planning algorithm based on TD3,TD3 algorithm and combined with the characteristics of the task based on the actual needs,determine the algorithm implementation plan,and targeted the design and use of the seven kinds of incentive function,especially the gripper at the end of the gesture constraints is decomposed into three function of reward,refining the hand claw and the specific situation of the target in the process of fetching,Effectively improve the success rate of training and testing;Finally,experimental tests were carried out in two scenarios,including system verification,model training and model testing,which verified the validity and scientificity of the research work in this paper.