【作者】 李明富；

【导师】 李世其；

【作者基本信息】 华中科技大学 ， 机械设计及理论， 2009， 博士

【摘要】 利用视觉信息控制机器人的运动，关键的难点在于视觉空间和机器人运动空间之间的非线性关系。目前人类设计的基于视觉控制的机器人系统无论在精度、稳定性还是鲁棒性上都无法和人类本身相比拟。人类(灵长类动物)通过其内部运动控制机制很好的解耦了视觉空间到运动空间之间的非线性关系，并且具有极好的鲁棒性。因此很多学者提出了多种仿生物机器人手眼协调系统，作出了开创性的研究。首先对神经生物学和人类运动学等领域关于视觉引导运动控制机制方面的研究进行了分析，从中提炼出可以为机器人领域所借鉴的一些内在机制。并根据这些内在机制构建了一个仿生物机器人手眼协调控制系统，该控制系统综合考虑生物视觉引导运动的固有特征和可变特征，将基于内部模型学习的视觉前馈控制和基于双目视差的视觉反馈控制引入到机器人控制系统中。针对机器人本体实时跟踪与感知问题，提出一种基于特征几何形状模板的、改进的CODENSATION跟踪算法；针对外部世界的目标物体，通过B样条曲线对其几何形状进行描述，并提出相应的CODENSATION跟踪算法；通过这两种不同的识别与跟踪策略很好的达到了机器人对其本体进行实时感知和对外部环境进行区分的目的。模仿生物视觉前馈控制的视觉-运动映射内部学习机制，提出了一种将BP神经网络和Sarsa(λ)强化学习相结合的学习算法，将这种学习算法应用到机器人视觉-运动映射模型中，将机器人姿态偏差角和视觉图像特征作为学习的输入量，将关节角速度作为学习的输出量，这种关节角速度可以解释为机器人的视觉前馈控制指令。仿真实验表明这种基于BP神经网络和强化学习的算法可以有效的逼近视觉-运动映射。针对传统的双目视差趋零控制存在的自由度限制和先验知识限制问题，提出了一种基于全局性轮廓特征和双目视差趋零控制的视觉反馈控制方法，构建由双目视差角误差和接近方向角误差组成的误差函数，从而推导出一种扩展的基于双目视差趋零控制的视觉反馈控制。这种扩展的视觉反馈控制可以较好的解决传统双目视差趋零控制中的自由度限制问题。同时目标跟随和接近实验表明这种扩展型视觉反馈控制可以有效的控制双目视差稳定趋零，从而达到精确对准和定位的目的。最后将手眼协调自主机器人应用到远程遥操作系统中构成监督控制遥操作系统和自主性可调的动态自主遥操作系统。为了和手眼协调自主机器人进行高效、流畅的交互，本文提出一种将自然图像界面和遥自主编程模块相结合的交互方式，通过这种交互方式将‘人—计算机—自主机器人’连接起来，自然图像交互界面为目标设定、定义提供渠道，遥自主编程模块为‘计算机—自主机器人’交互提供通道。实验表明这种应用手眼协调自主机器人的遥操作系统既可以发挥人类操作者的经验以应对较为复杂的工作环境，又可以充分的利用机器人的自主能力。更多还原

【Abstract】 The key difficulty of vision based robot control consists in the nonlinear relationbetween vision space and robotic motion space. Nowadays vision based robot controlsystem designed by humanbeings is not as good as itself in the aspect of precision,stability or robustness. The control mechanisms inherent in humanbeings (Primates)decoupling the nonlinear relation between vision space and motion space in highperformance with good robustness. And so, many reaseachers did the pioneer researchesand proposed kinds of robotic hand-eye coordination sytem which inspired biologically.Firstly, the accomplishment about the research of vision guided motion control inthe domain of neurobiology science and human body movement science is analyzed,mechanisms suitable for robotic control are abstracted from which. A framework ofbiologically inspired robotic hand-eye coordination is proposed based on thesemechanisms, the framework having considered both of invariant feature and variablefeature inherent in biological internal model. And the internal model learning basedvisual feedforward control and binocular disparity based visual feedback control isintegrated in the framework.Aimed to the problem of robotic body realtime tracking and percepting, animproved CONDENSATION tracking algorithm based on geometric shape template isproposed. As for external objects, the shape of objects is discribed by B-spline mothed,and relevant improved CONDENSATION tracking algorithm is proposed. By these twodifferent recognition and tracking motheds, the robot is endowed with the ability ofself-body real time perception and the ability of differentiation self-body to externalenvironment.A learning algorithm based on BP neural network and reinforcement learning isproposed by mimicing the biological visual feed-forward control which realized byvisual-motor mapping internal model learning. The algorithm is applied in roboticvisual-motor model learning, input of learning include robotic status deviated angle andimage feature in vision, output of learning is velocity of robotic joint, the velocity can beexplained to control command to realized the feed-forward control. The results of simulation experiment showed the algorithm based on BP neural network andreinforcement learning is effective to approximate the visual-motor mapping.As traditional binocular disparity to zero control is limited by degree of freedom andpriori-knowledge, a new visual feedback control based on roundly contour feature andbinocular disparity is proposed. In this mothed, the error function is consisted bybinocular error of vision and reaching angle error, a new expanded binocular disparityzero control based visual feedback control is deduced by the error function. Theexpanded visual feedback control can resolved the problem of DOF limitation whichconsisted in traditional binocular disparity control. Furthermore, experiment of objecttracking and reaching showed the expanded visual feedback control can control thebinocular disparity to zero stably, alignmenting and positioning accurately.Finally, the hand-eye coordination robot is applied to teleoperation. As a result, asupervisory control teleoperation system and a dynamic autonomy teleoperation systemwith adjustable level of autonomy were constructed by using the autonomous hand-eyecoordination robot as telerobot. For interact with autonomous hand-eye coordinationtelerobot, a new interaction composed by natural image interface and tele-autonomyprogram module is created. ’Human operator-computer-autonomous robot’ is connectedby the interaction, natural image interface provide a channel for object difinition andsetting, tele-autonomy program module provide tool for ’computer-autonomous robot’interaction. Experiments showed the teleoperation system based on autonomoushand-eye coordination robot not only exert human operator’s experience to treat complexwork environment, but also exert the robot’s capability of autonomy.更多还原