Online 2D-3D Radiation Mapping and Source Localization using Gaussian Processes with Mobile Ground Robots

Abstract

We present a novel method for online radiation mapping and source localization in 2D and 3D with mobile ground robots using Gaussian Processes to assist personnel in potentially dangerous scenarios such as nuclear catastrophes or dismantling nuclear reactors. While existing methods typically make strong model assumptions or are limited for robot onboard application by high computational cost, we propose a method that requires only weak model assumptions and gains efficiency by pre-sampling and local map update schemes. The resulting models can predict the radiation levels in complex indoor environments with multiple sources and quantify the uncertainty in their estimates. The proposed method can be applied in combination with teleoperated, semi-autonomous, or autonomous exploration. It was successfully evaluated at the EnRicH 2023 competition in a decommissioned nuclear power plant, where it provided the best localization and mapping of five radiation sources and received the award for radiation mapping. Our evaluation of data from the competition validates the accuracy and computational efficiency of the proposed approach. Moreover, we provide an open-source ROS implementation of the proposed method and open-access evaluation data.