Abstract & Overview
This paper proposes an Unscented Kalman Filter (UKF)-based IMU–LiDAR sensor fusion method for autonomous robot navigation in feature-poor enclosed spaces (homogeneous corridors, plain walls, changing lighting), conditions that often weaken LiDAR scan matching and cause large drifts in pure IMU integration. The proposed architecture models motion dynamics in error-state with online-estimated gyroscope/accelerometer biases, while LiDAR measurements are extracted as scan-to-submap constraints that are condensed into adaptive uncertainty relative pose observations. Time synchronization and extrinsic IMU–LiDAR calibration are performed on-the-fly using weak priors to maintain system stability despite time offsets. Evaluation on three indoor scenarios (a 40 m corridor, a 60 m L-aisle, and a 30 × 20 m warehouse with minimal texture) shows a 42–58% reduction in positional RMSE compared to LiDAR-only ICP and 73–81% compared to IMU-only, with translational drift < 0.6% of the distance traveled and heading drift < 0.35°/min. The system runs in real-time at 20–30 Hz on a mid-range CPU, maintaining a 100% localization success rate with no tracking failures at velocities of 0.4–1.2 m/s. These results confirm that UKF with adaptive uncertainty modeling and bias estimation is capable of integrating LiDAR inertial and geometric forces to produce accurate and robust state estimation in feature-poor indoor environments, while providing an efficient foundation for advanced trajectory planning and motion control.