Data-Driven, Ground Truth-Free Tuning of an Adaptive Monte Carlo Localization Method for Urban Scenarios

In this paper, we propose a tuning method for Adaptive Monte Carlo Localization (AMCL). The proposed method tunes the most important AMCL parameters without the need of a continuous ground truth by optimizing the estimated path smoothness and using the passage through a finite number of gateways as constraints. The optimization algorithm exploits Bayesian Optimization in order to limit the number of tuning runs.Data collected with an instrumented robot on a public road validate the approach. The proposed tuning yields a robust localization with minimal manual intervention in the tuning.