面向未知环境的理鞋机器人系统设计与实现

Design and implementation of shoes arrangement robot system for unknown environment

近年来机器人和人工智能技术发展迅速,但实现机器人在家居环境中的接触式操作仍然是一个具有挑战性的问题.为实现机器人自主整理鞋子的功能,提出利用实例分割网络和最小外接矩形识别鞋子和其朝向的方法,借助深度相机的点云信息,估计机器人的抓取位姿和放置位姿,并利用卷积神经网络和余弦相似度对同一双鞋子进行配对,设计了一套基于三维视觉的机器人自主理鞋系统.对系统中鞋子朝向识别的准确率和鞋子匹配的准确率进行评估,并进行真实机器人的自主整理测试.结果显示,所提方法对鞋子朝向识别的准确率达到96.2%,加入VGG16网络后,鞋子匹配算法的匹配准确率从62.6%提升到87.4%.该方法能准确实现机器人对鞋子及其朝向的识别进而对鞋子进行匹配,提升了机器人鞋子整理的稳定性.

In recent years,the robot and artificial intelligence technology have developed rapidly,but there still remains a challenging problem to realize the contact operation of robot in home environment.In order to make the robot arrange shoes automatically,we design an autonomous shoes arrangement robot system based on the 3D vision.In this system,the instance segmentation network and minimum enclosing rectangle are used to recognize the shoes and their orientation,and the grasping pose and placing pose of robot are estimated accurately by means of the point cloud information of depth camera.In addition,the convolution neural network and cosine similarity are adopted to match with a pair of shoes.Afterwards,we evaluate the accuracy of shoe orientation recognition and shoe matching in the system,and then carry out a real robot arrangement experiment.The result show that this method could assure 96.2%accuracy of shoe orientation recognition,and the matching accuracy of shoes increases from 62.6%to 87.4%when the VGG16 network is added to the shoes matching algorithm.In conclusion,this method can accurately recognize the shoes and their orientation,and then match with a pair of shoes,meanwhile,improves the stability of the robot shoes arrangement.