Research On Picking Tasks Method Of Nuclear Robot Based On Deep Reinforcement Learning

Radioactive waste picking tasks often face complex working environments such as unstructured and partially radioactive.With the flourish of robot control technology,manual picking methods have gradually been replaced by automated machines.The picking method using machinery remotely controlled by engineers has become mainstream.However,it also brings various challenges such as low picking efficiency,difficult operation,long personnel training cycle,and poor autonomous control,etc.Aiming at providing insights on the above problems,the thesis conducts comprehensive research on the method of picking tasks of nuclear robots based on deep reinforcement learning to improve the adaptability,efficiency and autonomous operation capability of nuclear robots in unstructured environments.First,the picking task characteristics of the nuclear robot are deeply analyzed in the thesis.In addition,the framework of nuclear robot picking operation system is designed,which consists of three parts: environment perception,data training and motion control.And the deployment mode between vision system and the robot is also discussed and selected.At the same time,the kinematics and inverse kinematics of the robot used in the system are analyzed.Combine the camera calibration principle to build a hand-eye calibration model to obtain the conversion relationship between image information and robot control.Then,a picking operation method of nuclear robot based on deep reinforcement learning is proposed,which is named as FR-DDQN.The traditional deep Q-learning network algorithm is improved,and the algorithm framework based on a double deep Q-learning network and prioritized experience replay is used to improve the training speed and stability of the algorithm.Based on the Markov decision process,the mathematical model of picking tasks is designed and constructed.The state space of image information and the action space of manipulator are built.The value function of a full convolution neural network is used to evaluate and output each pixel point in the input image information.An effective reward function is also designed with additional performance indexes on the basis of the radioactive area.It is to achieve the purpose of preferentially picking wastes with high radioactive activity.What's more,the method trains the grasping operation and the pushing operation respectively,and adopts a cooperative picking method to improve the picking efficiency and solve the problem of mutual influence between wastes.Finally,extensive experiments including simulation and physical verification were conducted.Multiple groups assigned to different tasks were performed to train and test of various types of radioactive solid waste picking to evaluate the performance of the proposed method.The experimental results show that the mechanical arm can independently complete the picking task under complex conditions,and the working efficiency can be improved by using the cooperative operation method of pushing and grasping.Furthermore,the mechanical arm will give priority to grasping objects with high radioactivity without being affected by waste stacking.In addition,the physical experiments verifies that the algorithm has migration ability and certain generalization.