【计算机科学】【2011】Any-Com多机器人路径规划

本文为美国科罗拉多大学（作者：Michael Wilson Otte）的博士论文，共194页。

在不太昂贵的方法失败的情况下，鲁棒的自主机器人系统必须使用完全或概率完备的路径规划算法（其中完备性是能够在解存在时找到问题解决方案的算法性质）。然而，这些方法有时计算复杂，无法在合理的时间内找到解决方案。机器人之间的通信倾向于增加完整性和降低计算复杂度；然而，通信质量是环境依赖的，并且经常超出用户或系统的控制。以前的多机器人路径规划问题的方法都是在完整性、计算性与通信性的状态空间中被裁剪成一个单一的点，并且通常不适合解决设计范围之外的问题。相比之下，我相信真正鲁棒的多机器人导航只能通过在该状态空间中自动调整其性能的算法来实现，以最大化性能相对于系统所面临的每个问题。

我个人偏向于最大化算法完备性，同时尊重当前可用的计算资源和通信质量。为了具备实用性，所得到的解必须在合理的时间内计算出来。我也相信，将寻找多机器人路径规划解决方案的计算工作拆分到所有机器人中是有意义的，因为该解决方案将从中受益。这可以通过重新组织一个联网的机器人团队作为一个ad-hoc分布式计算机来实现，从而允许团队的计算资源汇集起来，从而增加在特定时间内可解决问题的复杂性。然而，ad-hoc框架中的分布式计算必须考虑到计算节点（即机器人）之间的通信通常是不可靠的。我撰写的论文“共享任何时间的搜索进展，从一个动态机器人团队创建的自组织分布式计算机，使概率完备的、集中的、多机器人路径规划跨越广泛的复杂情况、通信质量和计算资源的难题。”

本文提出的支持我论文的工作可以分为三个相关的重点领域。

（1）提出了一种新的分布式规划概念，称之为随机嫁接搜索树耦合森林（C-FOREST），并证明了它在许多问题上的并行效率都大于1。

（2）提出了使用机器人团队作为自组织分布式计算集群，并证明当C-FOREST在这种类型的体系结构上运行时，它能够利用完美的通信，但随着通信质量的恶化，也会使得优良的性能下降。我用“Any-Com”这个词来描述具有后者特性的算法。

（3）提出了一个动态团队版本的Any-Com C-FOREST，它允许多个机器人团队形成合力，然后随着机器人在环境中移动而重新组织。每个团队充当一个ad-hoc分布式计算机来解决其复合机器人的公共路径规划问题。限制团队只包括冲突的机器人提高算法性能，因为它大大降低了每个团队必须解决的问题的计算复杂度。仅通过发生冲突的配置空间的子集重新规划具有类似的计算优势。

Robust autonomous robotic systems must usecomplete or probabilistically complete path planning algorithms when lessexpensive methods fail (where completeness is the algorithmic property of beingable to find a solution to a problem when one exists). However, these methodsare sometimes so computationally complex that a solution cannot be found withina reasonable amount of time. Communication between robots tends to increasecompleteness and reduce computational complexity; however, communicationquality is environmentally dependent and often beyond control of the user orsystem. Previous approaches to the multi-robot path planning problem have eachbeen tailored to a single point within the completeness vs. computational vs.communication state space, and are often ill-equipped to solve problems outsidetheir design envelope. In contrast, I believe that truly robust multi-robotnavigation can only be achieved by algorithms that automatically tune theirperformance within this state space to maximize performance vs. each problemthe system faces. My personal bias is to maximize algorithmic completenesswhile respecting the computational resource and communication quality that iscurrently available. In order to be useful, the resulting solutions must becalculated within a reasonable amount of time. I also believe that it makessense to divide the computational effort of finding multi-robot path planningsolutions among all robots that the solution will benefit. This can beaccomplished by recasting a networked team of robots as an ad-hoc distributedcomputer—allowing the team’s computational resources to be pooled increases thecomplexity of problems that can be solved within a particular amount of time.However, distributed computation in an ad-hoc framework must respect the factthat communication between computational nodes (i.e., robots) is usuallyunreliable. I propose the thesis “Sharing Any-Time search progress over anad-hoc distributed computer that is created from a dynamic team of robotsenables probabilistically complete, centralized, multirobot path-planningacross a broad class of instances with varied complexity, communicationquality, and computational resources.” The work presented in this dissertationin support of my thesis can be divided into three related areas of focus. (1) Ipropose a new distributed planning concept called Coupled Forests Of RandomEngrafting Search Trees (C-FOREST), and demonstrate that it has parallelizationefficiency greater than 1 for many problems. (2) I propose using a robotic teamas an ad-hoc distributed computing cluster, and demonstrate that when C-FORESTis run on this type of architecture it is able to exploit perfect communicationwhen it exists, but also has graceful performance declines as communicationquality deteriorates. I coin the term “Any-Com” to describe algorithms with thelatter property. (3) I propose a dynamic team version of Any-Com C-FOREST thatallows multiple robotic teams to form and then re-form as robots move about theenvironment. Each team acts as an ad-hoc distributed computer to solve itscomposite robots’ communal path planning problem. Limiting teams to includeonly conflicting robots improves algorithmic performance because itsignificantly reduces the computational complexity of the problem that eachteam must solve. Replanning through only the subset of the configuration spacein which conflicts occurs has similar computational benefits.

1. 引言
2. 相关工作
3. C-FOREST：超线性加速的分布式路径规划
4. Any-Com C-FOREST：基于网络机器人团队创建的ad-hoc分布式计算机路径规划
5. 动态团队Any-Com C-FOREST：具有动态团队的集中式多机器人路径规划
6. 结论
   附录A 路径规划背景知识
   附录B 任意时间最短路径随机树路径规划算法（任意时间SPRT）
   附录C 随机图中贪婪路径的期望长度

更多精彩文章请关注公众号：