AMLC-SLAM: adaptive multi-descriptor loop closure for LiDAR-inertial mapping in mobile field robots

Purpose This paper aims to improve long-term simultaneous localisation and mapping (SLAM) based on light detection and ranging (LiDAR) and an inertial measurement unit (IMU) for mobile field robots operating in large-scale, weakly structured and dynamically occluded environments, where a single loop-closure descriptor often cannot balance retrieval speed and geometric reliability. Design/methodology/approach AMLC-SLAM is built on a tightly coupled LiDAR–IMU mapping framework and introduces an adaptive multi-descriptor loop-closure module. A height-enhanced stable triangle descriptor (STD) strengthens local geometric discrimination, while an average-height ScanContext descriptor provides efficient global place retrieval. A distance-driven strategy selects lightweight radius search for nearby closures and activates ScanContext–STD candidate retrieval with voxelised generalised iterative closest point (VGICP) verification for long-distance revisits or large-drift cases. Verified loop constraints are inserted into a factor graph implemented using Georgia Tech Smoothing and Mapping (GTSAM) and incremental smoothing and mapping (iSAM2) for incremental global correction. Findings Experiments on KITTI, MulRan and a campus data set show that AMLC-SLAM improves the balance between loop-detection accuracy, localisation precision and runtime efficiency. The adaptive strategy achieves a 95.3% loop-detection success rate with an average processing time of 15.7 ms per frame. Compared with ScanContext-enhanced LiDAR–inertial odometry via smoothing and mapping (SC-LIO-SAM), the root mean square error (RMSE) is reduced by about 23.4% on MulRan KAIST 02 and 58.3% on CampusData 3. Research limitations/implications The method mainly relies on geometric descriptors, and performance may still degrade in extremely dynamic, severely occluded or geometrically degraded scenes. Practical implications The proposed method can improve long-term mapping reliability for mobile robots in campus, park, urban-road and suburban-road environments, especially under low-channel LiDAR sensing and repeated revisits. Originality/value The proposed method does not simply stack multiple descriptors; it uses the spatial relationship between keyframes to choose the loop-closure path. This provides an interpretable and deployable solution for robust mapping in campus-like, urban and suburban field-robot scenarios.