SuperLoc: The Key to Robust LiDAR-Inertial Localization Lies in Predicting Alignment Risks

TL;DR

SuperLoc tackles degraded-environment LiDAR localization by predicting alignment risk and estimating observability before optimization, enabling proactive, degeneration-aware sensor fusion. It introduces a covariance-based observability model and a pose-prior factor to weight auxiliary odometry where constraints are weak, integrated via a joint optimization framework. Empirical results show a 54% accuracy improvement over strong baselines and enhanced robustness across caves, multi-floor buildings, and long corridors, with online performance at 22 FPS. The work also provides open-source code and eight challenging datasets to advance research in robust map-based LiDAR localization.

Abstract

Map-based LiDAR localization, while widely used in autonomous systems, faces significant challenges in degraded environments due to lacking distinct geometric features. This paper introduces SuperLoc, a robust LiDAR localization package that addresses key limitations in existing methods. SuperLoc features a novel predictive alignment risk assessment technique, enabling early detection and mitigation of potential failures before optimization. This approach significantly improves performance in challenging scenarios such as corridors, tunnels, and caves. Unlike existing degeneracy mitigation algorithms that rely on post-optimization analysis and heuristic thresholds, SuperLoc evaluates the localizability of raw sensor measurements. Experimental results demonstrate significant performance improvements over state-of-the-art methods across various degraded environments. Our approach achieves a 54% increase in accuracy and exhibits the highest robustness. To facilitate further research, we release our implementation along with datasets from eight challenging scenarios

Figures (7)

• Figure 1: SuperLoc is an open-source LiDAR-inertial localization system that not only predicts alignment risks, and estimates the observability of scan, but also can actively incorporate pose priors from other odometry sources before failure occurs. The entire process doesn't require heuristic threshold adjustment to detect degeneration, and it has been evaluated in various challenging environments, including caves, long corridors, flat open areas, and staircases.
• Figure 2: System Overview of Proposed Method. The major contributions of this work are highlighted with a red square. The pipeline begins with LiDAR and IMU measurements, from which point, plane, and line features are extracted. Point-plane correspondences are used for PCA to determine principal and normal directions, which are analyzed for observability. Confidence values of each state direction are used to generate an observability scan, incorporating the pose priors into the sensor fusion system to prevent degeneracy.
• Figure 3: Real-Time Alignment Risk Analysis. From top to bottom: Cave, Multi-floor, and Long-corridor environments. Left: Observability scans, where red, green, and blue points represent accumulated observable features in the x, y, and z directions, respectively. Right: Histograms depicting real-time confidence values $[0, 1]$ for each direction. Lower confidence values indicate higher alignment risks.
• Figure 4: Localization in Cave Environments. Our method demonstrates superior performance with a significantly lower outlier rate of 0.50%, marking a 6-fold times improvement over the second-best method. The pink highlights at the bottom shows low confidence in the X direction, where our method actively fuses alternative odometry in X direction.
• Figure 5: Localization in MultiFloor Environments. Our method has significantly fewer outliers, with a rate of 8.03%, marking a 7.2-fold times improvement over the second-best method. The blue dashed square at the bottom highlights low confidence in the Z direction, where our method actively fuses alternative odometry in Z direction.