Research On Robot Grasping Based On Reinforcement Learning With Dynamic Motion Primitive

With the development of robot technology,bionic mobile robot combines machine vision,sensors information,neural network algorithm and control theory make the robot more intelligent.It can accomplish more complicated operation.However,the bionic robot doesn't have the ability of self-learning,and only receive control signals.So how to make the robot has the self-learning ability,making the robot to become an agent,which is a key point.The complex movements can be made up of a series of simple primitives action.The dynamic movement primitives is a method of path planning,and the aim of this work is to find a way to express the complex movements that can be adjusted flexibly without manual parameter tuning or having to worry about instability.Reinforcement Learning algorithm is a kind of machine learning algorithm that different form supervised learning.It doesn't need a priori knowledge,and improves self-learning and on-line learning ability through constant trial-and-error interaction with the environment.In this paper,the dynamic motion primitives algorithm combine with the Reinforcement Learning algorithm,so that the robot can automatically adjust the parameters of the dynamic motion primitives element to realize the self-adjustment of the trajectory.In this paper,the mobile robot platform is developed by laboratory which are used as experiment platform to simulate human grasping process.It consist of two manipulator s with five degrees of freedom(DoF),a bionic hand with six DoFs(left and right),and two wheels(two DoFs).The mobile robotic system can move to grasp object with the hand-eye system.The binocular stereo camera mounted on the robot can obtain the coordinate of the target object.The robotic system composed of two wheels and a 5-DoF manipulators,and can avoided the singularity of the operating space with the visual guidance.The mobile robot system is optimized by prim-dual neural network algorithm to obtain the trajectory of each joint.In this paper,we only take each joint position at the target point position.We get the target point as the target point of the dynamic movement primitive(DMP),which generate the desired curve,and tracking the curve.The Reinforcement Learning algorithm is applied to train the parameters of the DMP.After a number of learning,the bionic arm can be more precise reach the target location.The last step is the important part of the whole process.This time our robot has been able to reach the target object,through the force fingertips apply in the object to realize the tightly holding and fine manipulation.The recursive neural network algorithm is used to optimize the distribution of the force according to the position of the contact points.In this paper,the mobile robot can solve the redundancy problem in the operational space though the primal-dual neural network algorithm by visual guide.Then,the DMP algorithm generate the desired curve,and the Reinforcement Learning is used to train the training the DMP parameters,and achieve the target position.The recursive neural network algorithm is applied to optimize the grasping force,and finally complete the grasping task.