Methods Of Moving Object Grasp By Manipulator

The ability of robots to grasp objects like humans has always been one of the important goals of robot development.The robot's ability to capture moving objects is an essential function for applications such as automated production lines and on-orbit servicing.The problem of the manipulator grasping a moving obj ect is studied in this paper.The basic idea is to realize the stable tracking of the moving obj ect with the end effector of the manipulator at first,that is,to keep the position and posture of the end effector of the manipulator and the moving object within a certain range at a certain time.Then,the moving object can be grasped by the manipulator.In this paper,we focus on realizing the stable tracking of the moving object with the end effector of the manipulator.The velocity decomposition based planning algorithm and the eligibility trace based A3C(Asynchronous advantage actor-critic)algorithm are proposed to solve the problem of the moving obj ect grasping by the manipulator.The main work of this paper is detailed as follows:(1).The slow convergence speed of the traditional tracking algorithm always causes the grasping task failure because the moving object runs out of the working space of the manipulator before the grasping task is done.For this problem,this paper proposes the velocity decomposition based planning algorithm to realize the stable tracking of the moving object by the manipulator quickly.The success rate of the grasping of the moving objects by the manipulator is improved accordingly.The idea of this algorithm is to decompose the tracking problem into three sub-problems,and then solve each sub-problem with the optimal solution respectively to give the control command in Cartesian space,next,convert them into the control command in joint space with the model of the manipulator.With this control command,the manipulator can achieve stable tracking of the moving obj ects in a short period of time.On the third-party robot simulation platform,the simulation experiment of the manipulator grasping the moving object is carried out on the velocity decomposition based planning algorithm,and compared with a traditional tracking algorithm.The result verifies that the convergence speed of the velocity decomposition based planning algorithm is better than this traditional tracking algorithm.(2).For the velocity decomposition based planning algorithm,because the precise model of the manipulator is needed,the inaccuracy of the manipulator model will definitely lead to the performance degradation.A model-free reinforcement learning method called the eligibility trace based A3C algorithm is proposed in this paper to solve the problem of grasping the moving object by the manipulator.Because it is a method of learning through interaction with the environment,it is no longer necessary to know the model knowledge of the manipulator or the environment.The algorithm replaces the update of the A3C algorithm based on the n-step return with the update based on the eligibility trace.For the case of establishing the policy using Gaussian distribution in the continuous control problem and using entropy of the policy distribution to increase the exploration of the agent,the exploration will be mostly invalid due to the unlimited increase of the standard deviation,we propose to add the standard deviation constraints to increase the efficiency of exploration.Finally,through the established particle tracking simulation environment,the advantage of the sample utilization efficiency of the eligibility trace based A3C algorithm is verified to be better than the A3C algorithm.At the same time,the effectiveness of the proposed algorithm is verified by the established moving obj ect grasping simulation environment based on the third-party robot simulation platform.