【计算机科学】【2010.02】协作感知的多目标路径规划

本文为布拉格捷克理工大学（作者：Jan Faigl）的博士论文，共182页。

本文研究多目标路径规划问题，主要涉及搜救行动中协作机器人的探测任务。考虑了连续和离散两种传感模型。我们假设环境是先验已知的，在两种方法中都用多边形域来表示。与移动成本相比较，连续感知的成本相对便宜，这样可以建模为观察者路由问题（Watchman Route Problem，WRP）；离散感知将导致解耦方法的使用，其动机是感知成本占据主导的问题。解耦方法由两个问题组成：寻找最小感知位置集合的问题和多目标路径规划问题，以确定访问目标位置的路径。如果必须考虑额外的可见度约束，例如可见范围、入射角，则可以找到一组感知位置作为美术馆问题（Art Gallery Problem，AGP）或传感器布置问题的解决方案。多目标路径规划问题可以表示为旅行商问题（Traveling Salesman Problem，TSP），这样机器人需要穿越城市之间的路径。这三个问题（WRP、AGP、TSP）都是NP-hard问题，因此可以考虑用近似算法在合理的时间内进行求解。将多个机器人的协作表示为WRP的多机器人变体形式，即使用MinMax准则的TSP问题。

本文提出一种新的传感器布置算法，称为边界布置，用于确定可见范围受限的一组传感器位置。将该算法与基于凸多边形划分和随机双采样方案的算法进行了比较，所提出的算法在寻找位置的数量和所需的探测路径长度方面都优于这两种算法。

为了解决多目标路径规划问题，在多边形域W中考虑了TSP的自组织映射（SOM）。SOM自适应过程根据输入到神经网络的城市来修改神经元的权重。权重表示组成节点环的各个节点，而节点环表示TSP的解决方案。节点-城市路径和距离的计算由三种方法支持：基于W凸划分的测地线路径近似、导航函数和W的三角形网格。该算法在计算上是实用的，并且为GENIUS启发式算法提供了有竞争力的结果。

将基于SOM的导航函数算法应用于广义多目标路径规划问题，其中目标可以由一组点而不是单个点表示。在具有分段感知位置的检测规划中，以及在感兴趣凸区域的协作访问问题中，我们证明了所提出算法的有效性。

提出了一种在多边形域W中求解可见度范围受限的WRP问题的自适应算法，连续感知的计算由W的三角形网格和W的凸覆盖所支持。该程序还用于寻找MWRP变体的解决方案：独立巡逻路线和从公共仓库开始的封闭路线。

将考虑运动学和动力学约束的多目标路径规划问题视为基于所寻找路径的轨迹生成问题，此外，提出了一种新的运动规划器RRT-Pathext。该规划器能够求解多目标运动规划问题，并用于WRP的SOM算法中节点环的再生。对于满足运动学和动力学约束的一个或多个移动机器人，通过求解WRP作为探测轨迹。

The thesis deals with the multi-goal pathplanning problem. The problem is studied in the context of the inspection taskof cooperating robots in a search and rescue mission. Two models of the sensingare considered: continuous and discrete. An environment is a priori known andit is represented by a polygonal domain in both approaches. The continuoussensing assumes relatively inexpensive cost of sensing in comparison to thecost of motion and can be formulated as the Watchman Route Problem (WRP). Thediscrete sensing leads to the decoupled approach that is motivated by problemswhere the cost of sensing is dominant. The decoupled approach consists of twoproblems: the problem of finding minimal set of sensing locations, and themulti-goal path planning problem to find a path visiting the found locations.The set of sensing locations can be found as a solution of the Art GalleryProblem (AGP) or sensor placement problem if additional visibility constraintshave to be considered, e.g. visibility range, incident angle. The multigoalpath planning problem can be formulated as the Traveling Salesman Problem (TSP)in which paths between cities have to be traversable by the robot. All threeproblems (WRP, AGP, TSP) are known to be NP-hard, thus approximate algorithmsare considered to find solutions in a reasonable time. The cooperation ofseveral robots is formulated as the multi-robot variant of the WRP, resp. theTSP, with the MinMax criterion. A new sensor placement algorithm, called theBoundary Placement is proposed to find a set of sensing locations withrestricted visibility range. The algorithm is compared with approaches based onconvex polygon partitioning and the randomized dual sampling schema. Theproposed algorithm outperforms both algorithms in the number of found locationsand also in the required length of the inspection path. The Self-Organizing Map(SOM) for the TSP is considered in a polygonal domain W in order to solve themulti-goal path planning problem. The SOM adaptation procedure modifies weightsof neurons according to city presented to the neural network. The weightsrepresent nodes that are organized into a ring of nodes and the ring representsa solution of the TSP. Computations of node–city paths and distances aresupported by three approaches: an approximation of the geodesic path based onthe convex partition of W, navigation functions, and a triangular mesh of W.The proposed algorithm is computationally feasible and it provides competitiveresults to the GENIUS heuristic. The SOM based algorithm with navigationfunctions is applied to the generalized multi-goal path planning problem wherea goal can be represented by a set of points instead of a single point. Theproposed algorithms are demonstrated in the inspection planning with segmentsensing locations, and in the problem of cooperative visits of convex areas ofinterest. A new adaptation procedure is proposed to solve the WRP withrestricted visibility range in a polygonal domain W. The computation ofcontinuous sensing is supported by a triangular mesh of W and a convex cover ofW. The procedure is also used to find solutions of the MWRP variants:independent patrolling routes, and closed routes starting from a common depot.The problem of the multi-goal path planning with respect to kinematic andkinodynamic constraints is considered as a trajectory generation based on thefound path. Moreover a new motion planner called RRT–Pathext is proposed. Theplanner is able to find a solution of the multi-goal motion planning problemand it is used to regenerate a ring of nodes in the SOM based algorithm for theWRP. A solution of the WRP is found as an inspection trajectory for one orseveral mobile robots satisfying kinematic and kinodynamic constraints.

1 引言
2 相关工作
3 问题描述与本文研究目标
4 寻找感知位置
5 协作机器人的多目标路径规划
6 一般的多目标路径规划问题
7 连续感知的探测规划
8 轨迹生成的多目标路径规划
9 结论
附录A 算法实验——AGP
附录B 算法实验——TSP/MTSP
附录C 算法实验——WRP/MWRP

下载英文原文地址：

http://page5.dfpan.com/fs/9lc8j2d21329d160e87/

更多精彩文章请关注微信号：