Degeneracy-resilient LiDAR-inertial SLAM for autonomous tunnel inspection robots

Reliable localization and mapping remain challenging for autonomous tunnel inspection because satellite positioning is unavailable and repetitive geometric structures weaken LiDAR constraints. This work presents a tightly coupled LiDAR-inertial SLAM framework that combines intensity-assisted feature extraction, degeneracy handling, and factor-graph optimization with switchable constraints. Geometric and intensity cues are jointly used to strengthen scan matching, weakly constrained directions are identified to restrict unreliable LiDAR updates while retaining inertial predictions, and switchable constraints reduce the influence of erroneous loop closures during global optimization. The framework was evaluated on 20 sequences from the Kimera-Multi dataset, comprising 13 non-tunnel and 7 tunnel sequences, and in field scenes. On this common subset, the proposed method achieved an average absolute trajectory error (ATE) of 3.875 m, a mean rotation error (MRE) of 0.598°, and a mean map entropy (MME) of −6.431. Compared with GLIM, the best-performing baseline overall, these results correspond to reductions of 5.05 % in ATE and 4.78 % in MRE, together with an absolute decrease of 0.130 in MME. The average point-cloud registration time was 15.3 ms per frame, supporting real-time operation on the evaluated platform. Overall, this work provides a practical localization and mapping framework for autonomous robotic inspection in underground tunnels.