RA-LIO: Active Three-Stage Robust Frontend With Dual-Check Adaptive Mapping and Multilevel Vertical Constraints

Tightly coupled LiDAR-inertial odometer (LIO) often encounters robustness bottlenecks during violent motion due to two core issues: first, passive keyframe acceptance strategies fail to filter low-quality measurements introduced by abnormal motion; second, the absence of effective constraints leads to cumulative vertical drift on rough terrain. To address these challenges, this article introduces robustness-aware LIO (RA-LIO), a novel framework powered by an active three stage robust frontend with dual-check adaptive mapping and multilevel vertical constraints. First, in the adaptive initialization stage, a motion stability precheck generates a more conservative initial pose for unstable motions. Subsequently, in the constrained optimization stage, a multilevel vertical (Z-axis) constraint is integrated into the scan-to-map optimization process to actively suppress vertical errors. Finally, in the postoptimization validation stage, a postoptimization consistency check evaluates the optimized results. Only validated frames are accepted as new keyframes; otherwise, they are discarded, preventing map contamination. Experiments demonstrate that RA-LIO maintains high accuracy comparable to Lidar-inertial odometry via smoothing and mapping (LIO-SAM) on the KITTI benchmark. Crucially, under aggressive motion conditions featuring violent jolts and rapid rotations, the proposed robust frontend exhibits outstanding performance, constructing clear maps where LIO-SAM fails. These results demonstrate the effectiveness of the proposed framework in significantly enhancing LIO robustness in extremely dynamic environments.