Simulation-only training for real-world localization integrity monitoring

Reliable self-localization is a prerequisite for the safe operation of autonomous mobile robots. Particle-filter-based methods such as adaptive Monte Carlo localization (AMCL) may silently fail under dynamic disturbances, map ambiguities, or motion uncertainty. This paper presents a data-driven localization monitoring framework that detects localization failures using Light Detection and Ranging (LiDAR) scans, occupancy maps, and particle filter statistics. The monitoring model is trained exclusively in simulation using randomized environments, without real-world training data. Gradient-boosting models are employed to capture nonlinear consistency relations, and temporal feature aggregation is introduced to model failure dynamics over time. Experiments in simulation and on real robotic platforms demonstrate reliable failure detection and generalization to unseen environments and motion uncertainty levels. Cross-platform transfer to a different robot is evaluated and identified as a limitation. Simulation-only training with domain randomization achieves an F1 score of 0.86 on real-world data.

Content

Subjects

localization monitoring

particle filter

sim-toreal transfer

gradient boosting

autonomous mobile robots

DDC Class

004: Computer Sciences

629.892: Robot