Bayesian Multidimensional Scaling for Multi-Robot Localization

This paper presents a distance mapping-based multi-robot localization method, which works with incomplete data. We make three contributions. First, we propose the use of multi dimensional scaling (MDS) for multi-robot localization. Second, we formulate the problem to accommodate partial observations common in multi-robot settings. We solve the resulting optimization problem using ´scaling by majorizing a complicated function (SMACOF)´, a popular algorithm for iterative MDS. Third, we take advantage of the motion information of robots to help the optimization procedure. Three policies are compared at each time step: random, previous, and prediction (constructed by combining the previous pose estimates with motion information). Using extensive empirical results, we show that the initialization by the prediction method results in better performance in terms of both accuracy and speed when compared to the other two initialization techniques.