基于轨迹匹配的模仿学习在类人机器人运动行为中的研究

发布时间：2017-12-27 14:07

  本文关键词：基于轨迹匹配的模仿学习在类人机器人运动行为中的研究　出处：《北京工业大学》2016年硕士论文　论文类型：学位论文

  更多相关文章： 模仿学习 类人机器人 轨迹匹配 概率模型 动态系统

【摘要】：模仿是人与动物技能学习的一种方法,将模仿学习机制赋予机器人系统,使其具有类似人的运动技能学习行为,快速地实现复杂运动技能的获取,是机器人仿生研究的重点内容之一。本文基于生物模仿的仿生机制,构建了机器人模仿学习的框架,以该框架为指导,围绕基于轨迹匹配的模仿学习在类人机器人中的运动行为进行研究,具体如下:(1)基于概率模型的机器人模仿学习研究对基于概率模型的模仿学习算法进行研究,利用高斯混合模型(GMM)进行轨迹编码,学习示教行为的特征,通过高斯混合回归(GMR)进行泛化处理,实现行为再现。针对书写任务中运动轨迹较复杂的问题,引入概率模型的模仿学习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。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,泛化效果良好。(2)基于Kinect的Nao机器人动作模仿系统的开发与实现为避开复杂繁琐的底层运动控制,使机器人能够通过学习实现运动技能的获取,有效提高其智能性,将体态感知技术与仿人机器人NAO相结合,以机器人的模仿学习框架为指导,开发并实现了基于Kinect的Nao机器人动作模仿系统。利用Kinect体感摄像机的骨骼跟踪技术,采集示教者骨骼点信息,计算各骨骼向量间的夹角得到各关节角变化信息,将其作为示教数据,通过高斯混合模型对示教数据进行表征学习,经高斯混合回归泛化处理后,映射到Nao机器人中,实现动作的模仿。实验结果表明,Nao机器人能够进行实时和离线的动作模仿,运动轨迹平滑而稳定,动作模仿的效果较好。(3)机器人模仿学习的在线调整问题研究为使机器人能够在任务环境发生变化的情况下,根据任务参数的变化作出相应的动态调整,使其具备在线调整能力,完成预定任务,将高斯混合模型与动态系统法相结合,对机器人模仿学习的在线调整问题进行研究。将动态系统的在线调整能力与高斯混合模型(GMM)的复杂轨迹的编码能力相结合,使动态系统的参数学习问题转化为高斯混合回归问题(GMR),为动态系统法提供了一种概率形式的表述;引入参数化高斯混合模型,基于DS-GMR模仿学习方法,重点对目标位置发生变化的任务场景下的机器人在线调整问题进行了研究与仿真实现。仿真实验结果表明,该方法在一定程度上具备高斯混合模型的轨迹编码能力和动态系统的动态调整能力,当任务环境发生变化时,能够作出相应的调整,具备一定的在线调整能力,且轨迹匹配性能较好,泛化能力进一步增强。本文对基于轨迹匹配的模仿学习在类人机器人中的运动行为进行了研究,对目前存在的模仿学习方法进行了较为系统地分析、总结,并在已有方法的基础上进行了一定的优化和扩展,对模仿学习的研究及其在类人机器人运动行为中的应用具有一定的参考价值。
[Abstract]:Imitation is a way to learn human and animal skills. It imparts the imitation learning mechanism to the robot system, making it have similar human movement skills learning behavior, and quickly achieve the acquisition of complex motor skills, which is one of the key contents of robot bionic research. In this paper, based on the biological mechanism of bionic imitation, construct a framework for robot imitation learning, based on the framework as a guide, around the trajectory, imitation motion behavior in humanoid robot in the study, based on the specific as follows: (1) the robot imitation probability model for study of imitation learning algorithm based on probabilistic model based on the Gauss mixture model (GMM) trajectory encoding, learning characteristics of teaching behavior, by Gauss (GMR) mixed regression generalization, realize the behavior of reproduction. Aiming at the complexity of the trajectory in writing task, we introduce the imitation learning of probabilistic model to represent and generalize the writing trajectory, and then achieve the acquisition of robot's writing skills. The experimental results show that this method has good behavior coding ability and anti-interference ability, and can achieve continuous Chinese character writing. It can carry out multi task learning through the extension of GMM, and then achieve the writing of track non continuous Chinese characters. The generalization effect is good. (2) based on the development of Nao robot action Kinect imitation system and the realization of the bottom movement to avoid complicated control, the robot can obtain through learning exercise skills, improve their intelligence, body perception technology and humanoid robot NAO combined with imitation learning framework for robot guidance and development the simulation system of Nao robot movement based on Kinect. The skeletal tracking technique using the Kinect somatosensory camera, collecting demonstrator skeleton point information, calculate the angle between the skeletal vector between the joint angle change information, as the teaching data, through the Gauss mixture model to characterize the teaching data through learning, generalization to processing to Gauss after mixing, mapping to the Nao robot in the implementation of action imitation. The experimental results show that the Nao robot can simulate the action of real time and off-line, the trajectory is smooth and stable, and the effect of action imitates is better. (3) study on the on-line adjustment problem of robot imitation learning for the robot can in the task environment, make corresponding adjustment according to the change of task parameters, which has the ability to adjust online, is scheduled to complete the task, combining the Gauss mixture model and the dynamic system method, on-line adjustment of learning to imitate robot research. The dynamic system of the online adjustment ability and Gauss mixture model (GMM) complex locus encoding combined with the ability to make the parameters of the dynamic system of learning problem into the Gauss mixture regression problem (GMR), provides a form of probability expressions for the dynamic system method; introduced the parametric Gauss mixture model, DS-GMR imitation learning based on the method of research and Simulation on robot online adjustment of the target position changes of task scenarios focus. The simulation results show that this method has the ability to dynamically adjust the Gauss mixture model trajectory encoding ability and dynamic system to a certain extent, when the task environment changes, to make appropriate adjustments, have certain ability of adjustment, and the trajectory matching can further enhance the generalization ability. This paper studied the learning movement behavior in humanoid robot's trajectory matching based on the imitation of imitation, learning methods are systematically analyzed and summarized, and the optimization and expansion of some are based on the existing method, and its application in humanoid robot motion behavior has a certain the reference value for the study of learning.
【学位授予单位】：北京工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP242
，

本文编号：1342004