Reliability-Aware LiDAR Odometry Fusion for Mobile Robot Localization in Warehouse Environments

Abstract

Accurate self-localization is essential for reliable navigation of autonomous mobile robots in structured indoor environments such as warehouses and logistics facilities. In these settings, global positioning systems are unavailable, wheel odometry accumulates drift, and LiDAR scan matching can become unreliable when temporary obstacles or map changes create scan-to-map mismatch. Therefore, localization systems must combine the short-term continuity of odometry with the corrective value of map-referenced sensing while adapting to changes in sensor reliability.

This thesis investigates encoder- and LiDAR-based localization for a differential-drive warehouse robot using a progressive development and evaluation pipeline spanning MATLAB simulation and ROS 2/Gazebo implementation. First, a unified MATLAB framework is developed to compare four localization strategies under consistent conditions: encoder dead reckoning, LiDAR scan-to-map registration using Iterative Closest Point (ICP), Extended Kalman Filter (EKF) fusion of encoder prediction with ICP pose measurements, and particle-filter localization using a LiDAR likelihood model. The results show that encoder-only odometry is computationally efficient but accumulates drift, while scan-based and probabilistic methods can reduce long-term error by anchoring the pose estimate to environmental geometry. EKF fusion provides a lightweight and continuous estimate that combines encoder smoothness with LiDAR-based correction.

Second, the localization workflow is implemented in ROS 2 and evaluated in Gazebo under simulated-time operation, with RViz used for visualization. The implementation establishes a modular architecture for simulation, LiDAR preprocessing, local ICP odometry, map-based localization, EKF fusion, trust monitoring, and visualization. RTAB-Map is used for LiDAR-based local ICP odometry, AMCL is used for map-based probabilistic localization, and the robot_localization EKF fuses wheel odometry with ICP odometry to produce a smoothed /odometry/filtered estimate. A custom online ICP trust monitor provides a heuristic diagnostic score for local scan-matching quality. The ROS 2 results demonstrate the expected qualitative behavior of wheel odometry, local LiDAR odometry, and EKF fusion while also reflecting practical middleware effects such as asynchronous sensor updates, transform handling, and timing variation.

Third, this thesis introduces a reliability-aware localization enhancement for handling scan-to-map mismatch and accumulated odometry drift. When ICP becomes unreliable due to temporary obstacles or degraded scan matching, its influence is reduced to limit the effect of incorrect scan alignments on the estimate. When ICP remains consistently reliable, it is used as a geometric anchor to gradually reduce accumulated encoder drift. A Mamdani fuzzy inference system maps ICP registration RMSE, position innovation, and yaw innovation to a continuous trust score. Simulation results show that the proposed strategy maintains low estimation error during the tested map-mismatch condition and reduces encoder-ICP divergence during stable periods without abrupt switching. These results are limited to the simulated environments, trajectories, and mismatch conditions considered in this thesis.

Overall, the contributions of this thesis are: (i) a unified MATLAB evaluation framework for encoder-, ICP-, EKF-, and particle-filter localization in a representative warehouse environment; (ii) a reliability-aware localization strategy that adapts LiDAR trust based on scan-matching quality and innovation consistency; and (iii) a modular ROS 2/Gazebo implementation that integrates wheel odometry, local LiDAR odometry, AMCL, EKF fusion, and online ICP trust monitoring as a foundation for future hardware deployment. Future work will focus on physical robot validation, richer sensor-artifact modeling, systematic evaluation under a wider range of operating conditions, and full integration of the reliability-aware trust framework into the ROS 2 fusion loop.