Design Of Robot Behavior Control System Guided By Vision

At present,the world is facing an aging population,which has caused a series of severe social problems.As the family structure model tends to be miniaturized and the pace of life continues to accelerate,the issue of pensions has become particularly prominent,bringing severe challenges to the development of society.The emergence of intelligent robots provides a new way to solve the above problems.Based on the research of machine vision algorithms and robot behavior control algorithms,this thesis designs and implements a vision-guided robot behavior control system.The main work of this thesis includes:(1)Object recognition and localization based on depth camera.Firstly,this thesis took953 indoor scene images containing 44 kinds of objects,labeled them with Label Img,and formed the training set together with some images of the Microsoft common objects in context(MS COCO)data set and Pascal visual object classes(Pascal VOC)data set.Secondly,transfer learning is adopted to train the YOLOv4 algorithm-based target detection network.Finally,the depth value of the object to be grasped(target object)is obtained through the depth camera.According to Socket programming,the coordinate position of the target object is transferred to the lower computer to guide the robot’s behavior.The experiments show that,the target detection accuracy rate reaches 94.17% by loading the trained target detection network.(2)Robot behavior control based on back propagation(BP)neural network.According to the space coordinate position of the target object provided by the host computer,a BP neural network is built to fit the mapping relationship between the space position of the target object and the duty cycle of the pulse width modulation wave of each steering gear of the manipulator,and the network training and testing is completed.The experiments show that,the success rate of the robot arm grasping objects reaches 90.03% by loading the trained BP neural network with the pressure sensor.(3)Robot behavior control based on deep reinforcement learning network.The robot behavior control method based on BP neural network has some problems such as uncoordinated actions and invalid actions.Meanwhile,the pure BP neural network is an open-loop system.To solve the problems,this thesis uses deep reinforcement learning to control robot behavior.Traditional deep reinforcement learning only takes the distance between the end effector of the robotic arm and the target object as the reward function.This thesis presents a new reward mechanism which can increase the angle and relative distance constraints.The experiments show that,the proposed method can improve the problems of uncoordinated and invalid actions of the manipulator.After system debugging and testing,this system can complete the identification and positioning of a variety of objects in an indoor environment,and assist people with limited mobility to complete the object retrieval,which meets the system design requirements.This system has important theoretical significance and practical value for promoting the promotion and application of home service robots.