【作者】 薛颂东；

【导师】 曾建潮；

【作者基本信息】 兰州理工大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 群机器人是由数量众多、结构和功能相对简单的自主移动机器人组成的人工群体系统。在自组织机制下,通过有限感知和局部交互涌现群体智能,协同完成超出单个机器人能力的规定任务。其分布式协调控制策略源于生物群体的启发,具有鲁棒性强、柔性好、规模可伸缩等特点。本文针对目标搜索任务,通过建立反映生物群体特征的微粒群算法和目标搜索问题之间的映射关系,将前者扩展为群机器人的系统建模和协调控制工具,然后围绕定位、通信、避碰规划、信号融合等切实影响类微粒群算法发挥作用的主要环节,进行控制策略及其仿真研究。围绕这些内容所进行的工作及其成果如下：(1)就机器人与微粒的特征异同以及目标搜索与函数寻优的作用机理相似性,提出了系统层面上的机器人通信邻域结构和个体层面上的时变特征群概念,建立了微粒群算法与群机器人目标搜索问题之间的映射关系。在将机器人抽象为二维封闭空间中的运动微粒、将其动态特性抽象为一阶惯性环节后,给出了群机器人的扩展微粒群算法模型。在理想条件下基于该模型所做的目标搜索仿真实验,表明了该法的可行性,揭示了主要的算法参数与控制性能间的关系。(2)对于不便部署全局定位系统的搜索环境,提出了相对定位机制下扩展微粒群算法模型的形式化描述。首先由机器人的有限能力特性引入了短期记忆机制,根据机器人仅能记忆当前时刻和上一时刻的位置及相应信号的特点,确定了个体的最优认知位置和特征群的最优位置。由此,基于机器人的当前位置与个体认知位置和特征群最优位置间的相对观测距离和方位,确定了下一时刻个体坐标系下的期望速度和位置。仿真结果表明了该法的可行性,但也发现,系统的运行效率低于绝对定位机制下的目标搜索。(3)针对机器人之间的通信交互对控制算法效率的影响,提出了将个体认知和特征群共享信息分别更新的异步通信原则。机器人在各时间步采样后按扩展微粒群算法模型计算下一时刻的进化速度和位置,并随时更新自身认知,但对所记忆的关于特征群共享信息的更新则只在自身发现了更佳的特征群位置或监听到邻居机器人己对共享信息进行了更新时才进行。在此基础上,提出了基于通信周期和基于期望进化位置的异步通信策略。仿真结果表明,搜索算法中嵌入期望进化位置的异步通信策略后,系统具有较高的搜索效率。(4)针对机器人具有质量和物理尺寸,并受非完整运动学特性约束,且非结构化环境中存在障碍物和其他机器人的问题,提出了群机器人的避碰规划问题。首先基于机器人的多传感器结构,改进了人工势场法,并与扩展微粒群算法模型进行了有效集成。将各个时刻计算得到的一系列期望位置分别作为临时目标,对机器人产生吸引力,而接近传感器探测范围内的障碍物对其产生排斥力。合力决定了机器人的运动方向。而机器人在向临时目标移动过程中须服从自身的运动学特性约束。由此,确定了输入机器人控制器的控制向量,直至无碰撞地接近目标。仿真结果表明了该法的有效性。(5)针对位置评价以机器人对目标信号的检测为基础,且非结构化环境中目标信号具有多源异类等特点,故以一类典型的多源异类信号的传播环境为背景,提出了机器人对间歇性呼救声、周期性无线射频电磁波和连续性扩散瓦斯等实时性异类信号的融合框架。令目标和机器人分别作为信源和信宿,用单个信源和多个信宿之间各自独立的虚拟连续通信描述目标信号的检测过程。由此,给出了信源信号的三位二进制编码表。再引入二值逻辑,基于检测阈值提出了信宿侧的感知事件概念。然后,利用不同种类信号所具有的统计分布特性、可能籍此进行目标定位时的不同定位性质及精度,基于信息熵准则确定了不同种类信号对融合值的贡献度(权向量),对检测值进行归一化处理后作为信号的区分度,以此将异类信号融合为无量纲的纯量。在多源异类信号的传播环境中所进行的信号融合仿真,表明建立在该法基础上的目标信号搜索是可行的。更多还原

【Abstract】 As a novel branch of intelligent robotics, the focus of swarm robotics is on ap-plying swarm intelligence principles or techniques to the control of large groups of cooperating autonomous robots. In such a system, lots of homogenous robots hav-ing relative simple structure are controlled in a distributed fashion, being expected to emerge intelligence from limited sense and local interactions for accomplishing complex tasks. As to the task named target search, particle swarm optimization (PSO) is extended to model and control swarm robots. Being focused on this topic, some works are done and corresponding contributions are achieved in this thesis:(1) Due to the similarities and differences between particle and robot, swarm search can be mapped to PSO based on the common working mechanism.. To be-gin with, two concepts of communication neighborhood structure at systemic level and time varying characteristic swarm (TVCM) at individual level are introduced. Viewed as a particle moving in a closed two-dimensional workspace, each robot is abstracted as one order inertial element, and further, the model of extended PSO (EPSOM) can be given under an ideal environment. As for individual robots, each one is controlled by a three-state finite state machine.(2) In case of possible failure in deployment of global localization, relative po-sitioning mechanism is required. For this end, a short-time memory is introduced where each robot is capable of remembering signals measurement readings and po-sition coordinates at present and in the near previous time step. In this case, the cognitive position of each robot and the best-found position within its TVCM must be one of the two remembered positions. Then the relative observations including distance and bearing between the current position and the two best ones can be used to decide the expected velocity and position.(3) Local interactions among robots affect computation efficiency to a large ex-tent. Accordingly, an asynchronous parallel communication strategy is presented. In this case, virtual trigger with combinational logic is set for the shared social in-formation updating. The logical components are both finding signals fusion beyond the best social position by robot itself and monitoring update by other robots. Each robot updates its cognitive information at every time step, while its memory about social information updating occurs only when the trigger is sparked. If then, the robot will broadcast the updating information within its TVCM.(4) In the presence of obstacles and neighbors, path planning for each robot hav-ing physical size and kinematic properties have to be considered. Thus, a modified sensor-based artificial potential field (SBAPF) method and EPSOM is integrated for the problem solving after the kinematic model of individual robots are introduced. Each evolution position belonging to certain robot is taken as temporary target that attracts this robot. However, other robots and obstacles within its proximity sensors coverage repulse it. The moving direction depends on the composition of virtual forces. Besides, robot moving has to comply with its kinematic constraints.(5) PSO-type algorithm works on the premise that each robot fuses target sig-nals independently like fitness evaluate. Therefore, one typical environment where propagation of multi-source signals including intermittent call-for-help sound, peri-odic RF waves and continuous gas is brought forth. Let target and individual robot be virtual source and sink of information respectively. Then virtual continuous sink-source communication can be used to describe signals measurement. Because of the independence of heterogeneous signals emission from source, the signals can be encoded with three-bit codes at every time step and be transmitted to different sinks. Meanwhile, a concept of detection event can be proposed based on signals detection threshold at sinks. After then, those signals beyond threshold will be normalized and taken as distinguishing resolution. Also, some signal characteristics such as statistical properties, detection range, localization type and accuracy (if possible) can be used to decide the weights of contribution to fusion in dependence on comentropy-based criterion.All the above presented strategies are validated through a series of simulations conducted under satisfying corresponding conditions.更多还原