Intelligent Control Of Manipulator Based On Deep Reinforcement Learning

As artificial intelligence technology continues to advance,traditional manipulator applications are becoming increasingly intelligent.One of the key intelligence enhancements is to enable manipulator to gain the ability to quickly sort certain target items by physical attributes in an unstructured spatial environment.This paper uses the deep reinforcement learning algorithm DDPG(Deep Deterministic Policy Gradient)in manipulator control,instead of the inverse kinematic solving method,to control the manipulator to reach the target position through a data-driven training process,so as to grasp the target object and achieve the purpose of sorting target items.Firstly,different reward function forms and state vector forms are setup for training in a 2D manipulator simulation environment,comparing the convergence and convergence speed of the DDPG algorithm model and analyse the optimal setup.The DDPG algorithm was then migrated from a 2D simulation model to a 3D simulation model,and the DDPG algorithm was trained in the Gazebo 3D simulation environment of the ROS(Robot Operating System),where the optimal results were analyzed and compared,and the neural network weights were saved.Finally,a real Dobot Magician manipulator is controlled to grasp and sort the target items by using a visual recognition algorithm to identify and locate the target items and combining the saved neural network weights.This paper proposes a deep reinforcement learning-based training method for intelligent control of manipulator,which takes a 2D environment to 3D environment migration approach,greatly reducing the training time.The control model obtained by training is combined with a visual recognition algorithm to control the grasping and sorting operation of objects in the spatial environment by a real manipulator.The feasibility and effectiveness of applying reinforcement learning to real manipulator control is verified.