Synthetic Datasets for Data-Driven Localization Monitoring

Abstract

Reliable self-localization is fundamental to safe and efficient navigation in autonomous mobile robots and driverless industrial trucks. However, localization failures in highly dynamic or feature-poor environments can lead to safety hazards and costly workflow disruptions. While probabilistic methods such as particle filters mitigate sensing and actuation uncertainties, they lack mechanisms to recognize impending failures. To address this gap, we propose a systematic, physics-based simulation methodology for generating datasets that enable predictive failure detection. The datasets include localization estimates, ground-truth poses, sensor data, and automatically labeled failure cases. By systematically introducing challenging conditions, such as dynamic obstacles, featureless areas, and map ambiguities, we provoke diverse failure modes in a reproducible manner. These datasets establish a scalable foundation for training models that anticipate localization failures, supporting proactive fault detection and enhancing the safety and reliability of autonomous navigation in complex environments.