Abstract This study presents a LiDAR-inertial-visual odometry framework that maintains robust state estimation under perceptual degradation. Building on a tightly coupled error-state iterated Kalman filter, the method integrates modality-wise dual-confidence estimation with explicit degradation handling. A deep learning-based visual front end improves feature correspondence under illumination variations and texture-scarce scenes. During the measurement update, visual confidence is derived from the number of correspondences and reprojection error, while LiDAR confidence is estimated from point-to-plane residual statistics and the ratio of effective geometric features. These confidence scores are incorporated via dynamic covariance scaling to adaptively allocate modality contributions. When a modality becomes severely degraded, a degradation-aware transition policy regulates its influence through covariance inflation for soft suppression and temporary attenuation under persistent degradation, with a conservative recovery criterion to avoid frequent toggling. If a modality switch is required, poses are smoothly blended over a short transition window to preserve trajectory continuity and improve estimation stability. Experiments on NTU-VIRAL and Hilti’22 demonstrate competitive accuracy and consistent trajectories in degraded conditions, validating the robustness of the proposed framework.