基于深度强化学习微小软排线装配技术的研究  

Research on Micro-flexible Wire Assembly Technology Based on Deep Reinforcement Learning

作　　者：林杰 楚中毅[2] 任芸丹[3] LIN Jie;CHU Zhongyi;REN Yundan(Suzhou Luster VISION Intelligent Equipment Limited Liability Company,Suzhou Jiangsu 215000,China;School of Instrumentation and Optoelectronic Engineering,Beihang University,Beijing 100191,China;Suzhou Vocational University,Suzhou Jiangsu 215000,China)

机构地区：[1]苏州凌云视界智能设备有限责任公司,江苏苏州215000 [2]北京航空航天大学仪器科学与光电工程学院,北京100191 [3]苏州市职业大学,江苏苏州215000

出　　处：《机床与液压》2024年第14期89-93,共5页Machine Tool & Hydraulics

基　　金：科技创新2030“新一代人工智能”重大项目(2018AAA0102900)。

摘　　要：传统机器人控制方法仅限于固定种类和较为规则的来料,通过位置关系完成装配。由于排线的形态变异较大,很难实现抓取和自动化组装,排线的组装成功率和良率较低。针对宽度小于2 mm微小排线装配难题,通过机器3D视觉传感、力觉传感、触觉传感和本体觉传感等多模态融合技术,设计一套基于深度强化学习的微小软排线装配智能控制算法。在此基础上搭建了一组由协作机器人、六维力传感器、3D机器视觉系统组成的实验设备,并在多环境、不确定因素下验证了此方法装配的可行性。基于高精度微小排线的装配要求,通过深度强化学习多模态控制方法大幅提升了可靠性和装配的成功率,相比传统控制方法装配效率提升15%以上。此测试系统的装配精度可达±0.1 mm,装配成功率到达98%以上。Traditional robot control methods are limited to fixed types and relatively regular incoming materials,and the assembly is completed through the position relationship.Due to the large morphological variation of the wire,it is difficult to achieve grasping and automatic assembly,and the assembly success rate and yield of the wire are low.In order to solve the problem of micro-flexible wire assembly with a width less than 2 mm,a set of intelligent control algorithm for micro-flexible wire assembly based on deep reinforcement learning was designed by using multi-modal fusion technologies such as machine 3D vision sensing,force sensing,tactile sensing and proprioceptive sensing.On this basis,a set of experimental equipment compased of cooperative robot,6D force sensor and 3D machine vision system was built,and the assembly feasibility of this method was verified under multi-environment and uncertain factors.Based on the assembly requirements of high precision micro-flexible wire,the deep reinforcement learning multi-modal control method greatly improves the reliability and the success rate of assembly,while the assembly efficiency can be improved by more than 15%than traditional control condition.The assembly accuracy of this test system can reach±0.1 mm,and the assembly success rate can reach more than 98%.

关 键 词：机器人 深度强化学习 多模态融合技术 智能控制算法 

分 类 号：TP27[自动化与计算机技术—检测技术与自动化装置]