【作者】 王磊；

【导师】 孙增圻；

【作者基本信息】 清华大学 ， 计算机科学与技术， 2005， 博士

【摘要】 多机器人的合作是多机器人系统研究的核心问题。合作是机器人按照一定的原则组织在一起,通过完成各自分配的子任务来共同实现为了那些难以由单机器人完成的任务,所以机器人的组织关系和任务的分配是合作过程的主要困难所在。在动态对抗性环境下,根据对手的行为实时准确地给出对策,是对抗成功的关键,而对手意图识别是这一过程的核心问题。针对这些问题,本文提出了基于意图识别的多机器人动态联合合作方法。其中,建立了动态联合的多机器人系统体系结构,给出了任务分配的综合评价算法和对手意图识别的二次估计算法。动态联合是全局分布而局部集中以利益驱动的多机器人系统体系结构。动态联合结构保留集中方法能够最大限度的获得最优的任务解决方案的优点,采用局部弱集中,同时克服了集中方法的容错性能差的缺点;吸取了分布结构鲁棒性好的优点,又克服了其很难达到全局最优的缺点。利用市场模型的利益驱动机制,而摒弃了其复杂的成本计算和任务交易机制。任务分配的综合评价算法和一般的基于合同网的多边谈判不同,采用先通过综合评价值对机器人进行执行任务能力的评价之后,再和具有最高评价值的机器人进行一对一谈判。综合评价算法针对不同的全局最优目标,可以相应的调整评价体系的权重向量,以获得最佳的分配方案。首先以当前对手机器人的行为来估计其行为序列,再由行为序列估计意图,通过学习加强估计。该方法既避免了等待对手完整信息而失去及时性的缺点,又避免了单行为估计对手意图的不准确性。将对手意图识别的结果引入动态联合结构,对任务序列在四个层次上进行修正,作出对环境的动态反应,就实现了基于意图识别的多机器人动态联合合作。该方法在足球机器人系统中通过实际比赛得到了验证。更多还原

【Abstract】 The Cooperation is the key research problem of MRS (multi-robot system),which is that robots are organized by some rules to work together for thosetasks that cannot be completed by one single robot. The organization rulesform the architecture of MRS under which the tasks are decomposed intosubtasks and allocated to robots. So, how to construct architecture and how toallocate subtasks among robots are two difficulties must be solved forcooperation. The effective and exact opponent’s intention recognition isnecessary for MRS under dynamic environments to get predominance in acompetition. The real time reactive of robots depends on intentionrecognition.Architecture, task allocation and intention recognition are the problems thatthis paper tries to resolve by Dynamic Coalition method based on opponent’sintention recognition for MRS. This method consists of three parts ininterdependent relations. The first is an architecture named dynamic coalitionstructure for robots organization, the second is gift value algorithm for tasksallocation while the third is quadratic estimation algorithm for opponent’sintention recognition.Dynamic coalition structure is a profit-drive MRS architecture, which isglobal distributed while local weakly centralized. It can give optimal solutionof tasks as possible as centralized architecture because of local weaklycentralized coalition while get over the latter’s limitation of fault tolerance. Itadopts excellent robust performance of distributed architecture but presentsmuch more optimal tasks solving ability. It makes use of profit-drivemechanism like market model but avoids complex computation.Different from other common negotiation methods based on CNP, gift valuealgorithm for tasks allocation in this paper is two steps artifice. Firstly,estimate the task executing capabilities of all relative robots by gift value andselect the one who gives max value. Then, one to one negotiation happensbetween this robot and another robot, which is responsible for task allocation.The optimal solution of task allocation for different global objective can beobtained by adjusting weight vector for different estimating criterions in giftset.Quadratic estimation algorithm for intention recognition is also a two-stepartifice. The first step is to estimate the opponent’s behavior series and formsits current behavior. The next step is to recognize the opponents intentionfrom its behavior series gained in previous step. This algorithm overcomestwo difficulties. One is none-real time arising from waiting the totalinformation about opponent’s behaviors;the other is inaccurate arising fromrecognition by only single behavior.Import the opponents’ intention results in quadratic estimation algorithm, thenrevise task set to be executed in four layers, the reactivity of MRS toenvironments can be obtained. This is the work of dynamic coalitionarchitecture based on intention recognition.Finally, those methods have been applied and validated in robot soccercompetitions as described in the last chapter.更多还原