Global Localisation in Continuous Magnetic Vector Fields Using Gaussian Processes

Abstract

Localisation is one of the key capabilities for autonomous robots with sensors. Magnetic sensors to perceive the environment, although less explored, are an alternative modality to aid localisation. This paper proposes the use of continuous vector fields provided by a Gaussian Process (GP) with a divergence-free kernel that follows the magnetic flux to localise a mobile robot moving in a 2D space. The environment is pre-mapped in 3D producing a magnetic vector field with a GP that can be queried at any location. By means of a particle filter for Monte Carlo localisation, the mobile robot can be globally localised in the environment. We validate our approach using simulations and experimental results. Complex simulated environments using ANSYS are exploited to show our approach outperforms commonly used kernels.