返回
类型 应用研究 预答辩日期 2017-09-21
开始(开题)日期 2015-08-28 论文结束日期 2017-06-12
地点 中心楼315会议室 论文选题来源 中央、国家各部门项目     论文字数 8.27 (万字)
题目 移动侦测机器人的人机交互与局部自主关键技术研究
主题词 手势识别,走廊跟踪行为,避障行为,无标定视觉伺服
摘要 在战场环境侦察任务中或者发生恐怖袭击、核化泄漏事故时,派遣移动侦测机器人代替人类进入危险区域进行侦察、探测、取样或者应急处置,可以大大降低人员所受风险,并为决策指挥提供必要的信息支持。本文针对人-机器人交互和移动机器人局部自主两项关键技术开展相应的理论和实验研究。 首先根据移动侦测机器人工作环境,设计了一款小型移动侦测机器人,并根据移动侦测机器人半自主控制的需求,提出一种基于人机交互的机器人混合控制体系结构,该控制体系结构融合了基于认知模型的功能分解式体系结构和基于行为的反应式体系结构的特点。在移动机器人未见过的场景中或者在紧急情况下,人类操作者可以通过该控制体系对移动机器人的智能行为或基本行为进行直接干预,从而保证其应急处置的能力。 针对战场环境的特点,提出一种基于可穿戴式视觉的移动机器人手势控制系统,该手势控制系统包括手部检测分割和手势识别两部分。在手部检测分割部分,为了从可穿戴式视觉设备中采集到的包含复杂背景,剧烈光照变化等干扰的图像中提取出有效手部区域,提出一种基于轮廓线索和基于部件投票思想的手部检测器。在手势识别部分,通过将混合分类器和集成学习融合,提出一种深度集成混合分类器。为了减少连续手势切换过程中的中间态手势带来的误判,提出了一种投票滤波器,并将其应用于连续识别结果序列,该投票滤波器能够将连续手势切换过程中因为中间态手势带来的误判修正为期望的识别结果。 为了减少传统强化学习的试错次数并将其应用于真实场景的移动机器人的行为学习上,本文通过模拟人类学习过程提出一种模仿强化学习算法,该算法能够通过预先学习人类经验的方式提高移动机器人的学习速度并且显著减少试错次数。为了验证模仿强化学习的能力,本文将模仿强化学习应用于真实场景中的移动机器人的走廊跟踪任务中。实验部分验证了移动机器人在预先学习了人类经验的前提下只要再在走廊中经过很少次的试错学习后就能掌握走廊跟踪行为的控制策略。这说明本文提出的模仿强化学习方法确实可以使传统的强化学习方法不必局限于模拟仿真环境中,而是可以真正应用到真实场景的移动机器人的行为学习上。 为了使移动侦测机器人具备代替人类进入核污染环境进行放射源自主搜寻的能力,提出了一种小型移动机器人自主趋近核放射源算法,该算法使得移动机器人只需要一个2D激光雷达和两个核辐射剂量率仪就可以实现对单点核放射源的定位及自主趋近并且在自主趋近过程中进行自主避障。该算法包括趋核行为模块,陷阱逃离行为模块和行为仲裁模块,其中陷阱逃离行为是通过直接从人类的控制经验中学习而得到,人类经验的学习采用模糊进化神经网络。为了提高模糊进化神经网络的鲁棒性,降低人类操作经验中的错误经验对移动机器人带来的不利影响,采用一种混合进化算法对模糊进化神经网络的规则层的规则进行了优化,从而使得移动机器人可以不必拘泥于必须学习机器人控制专家的经验,即使普通人的控制经验仍然可以使移动机器人从中获取有效的控制规则。 为了使移动侦测机器人搭载的机械臂出现关节传感器故障时能够不召回地继续完成应急处置任务,提出一种适用于移动机械手无关节状态反馈情况的基于人-机-机协作的无标定视觉伺服控制系统。在辅助机器人的观察视角的帮助下,首先建立故障机械臂的虚拟外骨骼,虚拟外骨骼可以驱动故障机械臂,而虚拟外骨骼的末端使用人机交互方式产生的人工引导点引导。考虑到在引导过程中可能会存在因为非均匀引导而导致的瞬时大残差问题,提出一种残差切换算法,该算法能够根据人工引导点的运动特性自适应的更新关节角从而实现虚拟外骨骼末端对于人工引导点的稳定跟踪。为了实现虚拟外骨骼对于机械手的驱动,提出一种多关节模糊驱动控制器,它通过实时检测虚拟外骨骼各关节与机械手对应各关节之间的偏移矢量并通过最小化该偏移矢量来实现对机械手的驱动。实验部分验证了本文提出的控制系统能够直观方便的帮助操作人员实现对关节传感器出现故障的移动机械臂进行操作。
英文题目 RESEARCH ON CRITICAL TECHNOLOGY OF HUMAN-ROBOT INTERACTION AND LOCAL AUTONOMY FOR MOBILE RECONNAISSANCE ROBOT
英文主题词 Posture recognition, Corridor following behavior, Obstacle avoiding hehavior, Uncalibrated visual servoing
英文摘要 In battlefield reconnaissance missions, terrorist attacks or nuclear leakage accidents, sending mobile reconnaissance robot to replace human beings into the danger areas for reconnaissance, detection, sampling or emergency disposal, can greatly reduce the risk of personnel, and provide necessary information support for decision making. In this thesis, we carried out the theoretical and experimental research on two aspects of Human-Robot Interaction (HRI) and local autonomous behavior. Firstly, according to the working environment of the mobile reconnaissance robot, we de-sign a small mobile reconnaissance robot, and according to the needs of the semi-autonomous control, we propose a HRI-based hybrid robot architecture. The architecture combines features of the cognitive model based functional decomposition architecture and behavior based reac-tive architecture. In the unseen scene or in the emergency case, the human operator can control the mobile robot directly through the intelligent behavior or the basic behavior of the mobile robot, so as to ensure the mobile robot has the capacity of emergency disposal. According to the characteristics of the battlefield environment, an egocentric vision based posture control system is proposed. The posture control system consists of a hand detector and a posture recognition. In the hand detection and segmentation part, to extract the effective hand regions from the images that captured from the egocentric vision equipment and con-taining complicated backgrounds, large egomotions and extreme transitions in lighting, we propose a novel hand detector based on the contour cues and part-based voting idea. In the posture recognition part, a deep ensemble hybrid classifier is proposed by combing hybrid classifier and ensemble learning technique. To reduce misjudgments during consecutive pos-ture switches, a vote filter is proposed and applied to the sequence of the recognition results. This vote filter can correct the misjudgments results to the expected results. In order to reduce the number of trial-and-error in traditional reinforcement learning, and apply it to the behavior learning of mobile robots in real scenes, by simulating the human learning process we present an imitating reinforcement learning algorithm, the algorithm can enable the mobile robot to improve the learning speed and reduce the number of trial-and-error by learning the human operation experience in advance. To verify the performance of the proposed imitation reinforcement learning algorithm, we apply this algorithm to the corridor following task in real scence. The experiments show that the mobile robot can grasp the control strategy of the corridor tracking behavior only after the trial and error in the corridor under the premise of learning the human experience. This shows that our proposed imitating reinforcement learning method can make the traditional reinforcement learning method need not be limited to the simulation environment but can be transplanted to the behavior learning of mobile robots in real scenes. In order to make the mobile robot have the ability to replace human beings into the nuclear radiation environment for the pollution source searching task, we propose an nuclear radiation source autonomous approach algorithm for small mobile robot, the algorithm only requires a 2D laser radar and two nuclear radiation dose rate meter, and can realize the positioning of single point nuclear radiation source and autonomous approaching, and can avoid the normal obstacle in the autonomous approaching process. The algorithm consists of nucletaxis behavior module, trap escape behavior module and behavior arbitration module. The nucletaxis behav-ior can provide the guidance for the mobile robot according to the indicating value of the two nuclear radiation dose rate meter, and can avoid the normal obstacle for the mobile robot by the 2D laser radar. When the behavior arbitration module determines that the mobile robot has been trapped in the trap, the trap escape behavior module is activated. The trap escape behav-ior is obtained by learning directly from the human operation experience, and the fuzzy neural network is used to learn the human experience. In order to improve the robustness of the fuzzy neural network and reduce the adverse effects of bad human experience on a mobile robot, we propose a hybrid evolutionary algorithm for the optimizing of the rule layer of the fuzzy neural network, so that the mobile robot can not be confined to learn the robot experts’ experience, even ordinary people’s experience can still make the mobile robot learn the effec-tive control strategy. In order to ensure the mobile robot to continue the target disposal task without being re-called when the joint sensors (one or more) of the mobile robotic manipulator are faulty, we presents a HRRC-based (Human-Robot-Robot-Cooperation) uncalibrated visual servoing control system. With the aid of the surveillance video of the auxiliary robot, the virtual exo-skeleton of the faulty manipulator is built online, the virtual exoskeleton can drive the manip-ulator and the terminal of the virtual exoskeleton can be guided by the artificial guiding point which is produced by the HRI devices. Considering the instantaneous large residual caused by non-uniform guiding in the guiding process, we propose a Residual Switching Algorithm. It can update the joint angle formula according to the motion characteristics of the artificial guiding point, so as to ensure the tracking stability. To further drive the manipulator, we pro-pose a Multi-joint Fuzzy Driving Controller, which can drive the corresponding joint of the manipulator according to a offset vector between the virtual exoskeleton and the manipulator. The experiment shows that our proposed control system can assist the operator to control the mobile robotic manipulator intuitively, effectively and efficiently with faulty joint sensors.
学术讨论
主办单位时间地点报告人报告主题
东南大学仪器科学与工程学院 2013.11.26 中心楼304 纪鹏 机器人技术对工业及就业的影响
东南大学仪器科学与工程学院 2014.12.08 中心楼313 吴常铖 一种上肢康复训练机器人
东南大学自动化学院 2015.05.21 中心楼教育部重点实验室 Y. Raymond Fu When data meet uncertainty
东南大学仪器科学与工程学院 2015.09.03 中心楼304 纪鹏 基于最大互信息量和简化脉冲耦合神经网络的自适应图像分割方法
东南大学仪器科学与工程学院 2013.11.15 中心楼313 钱夔 发育神经网络原理及应用
东南大学仪器科学与工程学院 2013.10.9 中心楼304 纪鹏 多自由度仿人形上假肢
东南大学仪器科学与工程学院 2013.11.17 中心楼313 纪鹏 Application of wavelet transform in fingerprint identification
东南大学仪器科学与工程学院 2014.05.08 中心楼304 纪鹏 Accuate 3D shape measurement of multiple separate objects with stereo vision
     
学术会议
会议名称时间地点本人报告本人报告题目
International Symposium on Infrared Technology 2016.05.10 北京 Vision-based Posture Recognition Using an Ensemble Classifier and a Vote Filter
第三届国际质量·安全·信用论坛会议 2014.10.10 南京 Patrol Inspection Robot for Nuclear Power Plant
International Conference on Control Engineering and Artificial Intelligence 2017.01.14 吉隆坡 Self-adaptive Correction of Heading Direction in Stair Climbing for Tracked Mobile Robots Using Visual Servoing Approach
     
代表作
论文名称
Egocentric-Vision based Hand Posture Control System for Reconnaissance Robots
Vision-based Posture Recognition Using an Ensemble Classifier and a Vote Filter
适用于移动机械手无关节状态反馈情况的基于 人-机-机协作的无标定视觉伺服控制
Self-Adaptive Correction of Heading Direction in Stair Climbing for Tracked Mobile Robots Using Visu
 
答辩委员会组成信息
姓名职称导师类别工作单位是否主席备注
孙富春 正高 教授 博导 清华大学
刘文波 正高 教授 博导 南京航空航天大学
费树岷 正高 教授 博导 东南大学
陈熙源 正高 教授 博导 东南大学
宋光明 正高 教授 博导 东南大学
      
答辩秘书信息
姓名职称工作单位备注
曾洪 副高 副教授 东南大学