在书写任务中的基于轨迹匹配的模仿学习  

Imitation Learning Based on Trajectory Matching in the Writing Task

作　　者：于建均[1] 门玉森 阮晓钢[1] 徐骢驰 YU Jianjun;MEN Yusen;RUAN Xiaogang;XU Congchi(College of Electronic and Control Engineering,Beijing University of Technology, Beijing 100124, China)

机构地区：[1]北京工业大学电子信息与控制工程学院,北京100124

出　　处：《北京工业大学学报》2016年第8期1144-1152,共9页Journal of Beijing University of Technology

基　　金：国家自然科学基金资助项目(61375086);高等学校博士学科点专项科研基金资助课题(20101103110007)

摘　　要：针对书写任务中运动轨迹较复杂的问题,引入基于轨迹匹配的模仿学习算法对书写轨迹进行表征和泛化,进而实现机器人书写技能的获取.机器人从示教者处获取示教数据,利用高斯混合模型(Gaussian mixture model,GMM)进行编码,学习示教行为的本质特征,通过高斯混合回归进行泛化处理,实现行为再现.实验结果表明:该方法具有良好的行为编码能力和抗干扰性,能够实现轨迹可连续的汉字书写,通过对GMM的扩展能够进行多任务学习,进而实现轨迹不可连续汉字的书写,泛化效果较好.Aiming at the complexity of motion trajectory in the writing task, imitation learning based ontrajectory matching was introduced to represent and generalize writing trajectory for the obtainment of writingskill. The robot acquires training data from a teacher and codes by Gaussian mixture model (GMM). Then,the essential feature of teaching behavior was learned and the motion trajectory was reconstructed by means ofgeneralized output through Gaussian mixture regression (GMR). The results of simulation experiments showthat this method possesses favourable ability of behavior coding and anti-interference performance. It can beused in the writing task of Chinese characters whose trajectory is continuous. Furthermore, the learning ofChinese characters爷writing skill with intermittent trajectory is realizable by multi-task learning extended fromGMM and the generalization performance turns out to be good.

关 键 词：机器人 模仿学习 书写任务 高斯混合模型 高斯混合回归 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]