Range sensor-based robot localization using neural network

Localization is one of the most important requirements for mobile robot technology. Although many solutions have been proposed, there is no universally accepted solution. Most of the systems have some common drawbacks such as large computation power, expensive sensors, the scalability of environments, hard implementation, and the complexity of the system. The work presented in this paper deals with localization using ultrasonic range sensors and neural network. The proposed localization method can be implemented easily with several low cost range sensors and a low cost PC, thanks to the simplicity of neural network structure. Although neural network is trained in general environment for once, it is unnecessary to learn again in new environment. The experimental result shows that odometric errors are corrected with the proposed method.

1. INTRODUCTION

One of the main issues for a mobile robot is to determine and to keep track of its pose. Because dead reckoning error is unbounded, the use of sensor data which includes the information of the current environment must be necessary. Localization process can be changed into matching problem between two sets of sensor data.