Estimating the Odometry Error of a Mobile Robot by Neural Networks

Abstract

Localization is the accurate estimation of robot's current position and is critical for map building. Odometry modeling is one of the main approaches to solving the localization problem, the other being a sensor based correspondence solver. Currently few robot positioning systems support calibration of odometry errors in both feature rich indoor and landmark poor outdoor environments. To achieve good performance in various environments, the mobile robot has to be able to learn to localize in unknown environments, and reuse previously computed environment specific localization models. This paper presents a method combining the standard Back-Propagation technique and a feed-forward neural network model for odometry calibration for both synchronous and differential drive mobile robots. This novel method is compared with a generic localization module and an optimization based approach, and found to minimize odometry error because of its nonlinear input-output mapping ability. Experimental results demonstrate that the neural network approach incorporating Bayesian Regularization provides improved performance and relaxes constraints in the UMBmark method.