基于深度强化学习的密集物体温度优先推抓方法  

Temperature priority pushing and grasping method for dense objects based on deep reinforcement learning

作　　者：陈满 李茂军[1] 胡建文[1] 赖志强 李俊日 CHEN Man;LI Maojun;HU Jianwen;LAI Zhiqiang;LI Junri(School of Electrical and Information Engineering,Changsha University of Science and Technology,Changsha 410114,China)

机构地区：[1]长沙理工大学电气与信息工程学院,湖南长沙410114

出　　处：《传感器与微系统》2022年第1期41-44,49,共5页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(61601061);长沙理工大学研究生实践创新与创业能力提升项目(SJCX202051)。

摘　　要：对于工业中需要考虑温度特性的密集物体抓取时出现的操作难度系数大、温度优先性不强等典型问题,提出一种基于深度强化学习(DRL)的密集物体温度优先推抓(TPG)方法。方法涉及一种深度Q网络算法,以RGB-D图像与红外图像作为输入,使用两个全卷积网络(FCN),将推动和抓取放在一个框架内联合动作,输出对应的Q值,指导机器人对密集物体进行推抓动作,并与环境交互获得奖励。实验结果表明:对于所设抓取对象的完成率为100%;抓取成功率为69.4%,高于无推动的对比实验11.2%,温度相关度为79.5%,具有优先抓取温度较高物体的功能。To deal with the typical problems such as difficult to operate and low temperature priority when capturing dense objects that need to consider temperature characteristics in industry,a high-temperature priority pushing and grasping method for dense objects(TPG)based on deep reinforcement learning(DRL)is proposed.This method involves a deep Q network algorithm that takes RGB-D images and infrared images as inputs,uses two fully convolutional networks,puts the push and grab in a frame and outputs the corresponding Q value to guide the robot.And interact with the environment to get rewards.The experimental results show that the completion rate of the set grabbing objects is 100%,the grabbing success rate is 69.4%,which is higher than the non-driven comparison experiment of 11.2%,and the temperature correlation is 79.5%,which has the function of priority to grab objects with higher temperatures.

关 键 词：深度强化学习 全卷积网络 密集物体抓取红外图像 温度奖励 

分 类 号：TP241.2[自动化与计算机技术—检测技术与自动化装置]