Decoupled soft SVD Jacobian remapping for degeneracy-aware LiDAR-inertial SLAM

Abstract

LiDAR-inertial odometry (LIO) is easily degraded in geometrically degenerate scenes such as tunnels, subway corridors, and long roadside passages, where the observation Jacobian becomes ill-conditioned and the filter over-trusts poorly constrained directions. Existing SVD-based Jacobian remapping suppresses degenerate directions by hard truncating small singular values, but the hard cut-off causes update discontinuity, couples rotation and translation in one decomposition, and depends on a fixed threshold that is difficult to transfer across scenes. This paper presents a compact degeneracy-aware back-end update for FAST-LIO2. The method replaces hard truncation with a sigmoid soft weight, splits the pose Jacobian into rotation and translation blocks, and assigns each block an adaptive threshold proportional to its dominant singular value. Experiments on three WHU-Helmet sequences show that the proposed method consistently reduces vertical drift against FAST-LIO2 and DALI-LIO while maintaining real-time speed