基于结构文法的机器人模仿学习方法  

Robot imitation learning method based on structural grammar

作　　者：丛明[1] 简继盼 邹强 刘冬[1,2] CONG Ming;JIAN Jipan;ZOU Qiang;LIU Dong(School of Mechanical Engineering,Dalian University of Technology,Dalian 116024,Liaoning China;Dalian University of Technology Jiangsu Research Institute Co.Ltd.,Changzhou 213164,Jiangsu China)

机构地区：[1]大连理工大学机械工程学院,辽宁大连116024 [2]大连理工江苏研究院有限公司,江苏常州213164

出　　处：《华中科技大学学报（自然科学版）》2021年第10期97-102,共6页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(61873045);江苏省自然科学基金资助项目(BK20180190);大连市科技创新基金资助项目(2019J12GX043);中央高校基本科研业务费专项资金资助项目(DUT19JC56)。

摘　　要：针对机器人模仿学习方法泛化性弱、对低级检测器准确率要求较高的问题,提出了一种基于结构文法的模仿学习方法.该方法通过视觉传感器提取场景的符号描述,形成含有噪声的符号基元序列;采用概率上下文无关文法(PCFG)对这些序列进行表征和语法操作,从而形成语法空间;基于最小描述长度(MDL)准则对语法空间中的语法质量进行评价,用改进的Beam Search算法寻找最优语法,即演示活动的一般结构;获得的一般结构能够将含有噪声的符号基元序列进行解析,得到正确的序列.数据合成实验和汉诺塔实验的对比效果验证了该方法优秀的数据表达性能和良好的抗干扰性能,在高噪声环境下该方法的解析成功率约为90%.Aim at the problems of weak generalization of robot imitation learning method and high accuracy requirement of low level detector,an imitation learning method was proposed based on structural grammar.The method extracted the symbol description of the scene through the vision sensor,and then formed the symbol primitive sequence containing the noise.The probabilistic context-free grammar(PCFG)was used to characterize and manipulate these sequences,and the grammar space was formed.The minimum description length(MDL)criteria was used to evaluate the quality of the grammar in the grammar space,and the improved Beam Search algorithm was used to find the optimal grammar,which was the general structure of the demonstration activity.The general structure obtained could parse the symbol primitive sequence containing noise,and obtain the correct sequence.The excellent data expression performance and good anti-interference performance of this method was verified by comparing experiment effect of data synthesis experiment and Hanoi Tower experiment,and the parsing success rate of this method in high noise environment is about 90%.

关 键 词：机器人 模仿学习 结构文法 概率上下文无关文法 最小描述长度准则 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]