A Novel Method to Determine a Robot's Position Based on Machine Learning Strategies

摘要

An open problem in robotics is the one dealing with the way a mobile robot locates itself inside a specific area. The problem itself is vital for the robot to correctly achieve its goals. There are several ways to approach this problem, for example, robot localization using landmarks [4], [5], calculation of the robot's position based on the distance it has covered [6], [7], etc. Many of these solutions imply the use of active sensors in the robot to calculate a distance or notice a landmark. However, there is a solution which has not been explored and is the main topic of this paper. In essence the solution we tested has to do with the possibility that the robot can determine its own position at any time using only a single sensor, and a reduced database. This database contains all the information needed to match what the robot is sensing with its spatial position. In order for the method to be practically implementable we reduced the number of necessary matches by defining a subset of the original database images. There are two issues which have to be solved in order to implement such solution: a) the number of elements in every subset of the matching images and b) the absolute positions of each of these elements. Once these are determined, the matching process is very fast and ensures the adequate identification of the robot's position without errors. However, the two goals we just mentioned impose conflicting optimization goals. On the one hand we seek for the largest subset so that position identification is accurate. On the other we wish this subset to be as small as possible so that the online processing is as fast as possible. These conditions constitute a multi-objective optimization problem. To solve it we used a Multi Objective Genetic Algorithm (MOGA) which minimizes the number of pixels required by the robot to identify an image. To test the validity of this approach we also solved this problem using a statistical methodology which solves problem (a) --and a random mutation Hill Climber to solve problem (b).