D²-LIO: Enhanced Optimization for LiDAR-IMU Odometry Considering Directional Degeneracy
Guodong Yao, Hao Wang, Qing Chang · 2025
Read D²-LIO: Enhanced Optimization for LiDAR-IMU Odometry Considering Directional Degeneracy by Guodong Yao, Hao Wang, Qing Chang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-14355.txt.
Introduction
Guodong Yao, Hao Wang, and Qing Chang improve a navigation method that fuses laser scans with motion sensors to track a moving platform. Such systems map surroundings and locate the device, but they falter where laser features run thin — long corridors, open fields — because matches become unreliable. Their framework attacks that weakness in three ways.
First, a matching step drops bad point pairs using a threshold that adapts to distance and to how hard the platform is moving. Second, each new scan is registered against a local sub-map, with motion sensor data guiding the fit in weak-geometry spots. Third, a weighting scheme blends the sensor uncertainty with a measure of how degenerate the scene is. Tests indoors and outdoors, in sparse and degenerate settings, show the method beating leading rivals on toughness and accuracy. The work targets robots and vehicles that must not lose their place.
Catalog record
| CREATOR | Guodong Yao, Hao Wang, Qing Chang |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 33 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-14355
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
D²-LIO: Enhanced Optimization for LiDAR-IMU Odometry Considering Directional Degeneracy Guodong Yao, Hao Wang, Qing Chang Abstract LiDAR-inertial odometry (LIO) plays a vital role in achieving accurate localization and mapping, especially in complex environments. However, the presence of LiDAR feature degeneracy poses a major challenge to reliable state estimation. To overcome this issue, we propose an enhanced LIO framework that integrates adaptive outlier-tolerant correspondence with a scan-to-submap registration strategy. The core contribution lies in an adaptive outlier removal threshold, which dynamically adjusts based on point-to-sensor distance and the platform’s motion amplitude. This mechanism improves the robustness of feature matching in varying conditions. Moreover, we introduce a flexible scan-to-submap registration method that leverages IMU data to refine pose estimation, particularly in degenerate geometric configurations. To further enhance localization accuracy, we design a novel weighting matrix that fuses IMU preintegration covariance with a degeneration metric derived from the scan-to-submap process. Extensive experiments conducted in both indoor and outdoor environments—characterized by sparse or degenerate features—demonstrate that our method consistently outperforms state-of-the-art approaches in terms of both robustness and accuracy. I INTRODUCTION Recent advancements in robotics have substantially enhanced autonomous navigation and mapping capabilities, primarily driven by innovations in sensing technologies and algorithmic development [ 1 ] . Among these, simultaneous localization and mapping (SLAM) has emerged as a cornerstone technology, enabling real-time localization and map-building in unknown environments [ 2 , 3 , 4 , 5 , 6 ] . SLAM plays a crucial role across a wide spectrum of applications, including autonomous vehicles, aerial drones, and planetary rovers operating in global positioning system (GPS)-denied or complex terrains [ 7 ] . Among various sensor modalities, light detection and ranging (LiDAR) has gained prominence for its ability to capture high-precision 3D spatial information. LiDAR-based SLAM systems excel in generating detailed maps, enabling robust navigation in both structured and expansive outdoor environments [ 2 , 3 , 4 , 5 , 6 ] . However, despite their advantages, LiDAR-based systems face significant challenges in environments that lack distinctive geometric features—such as tunnels, forests, and open fields—where SLAM performance may degrade [ 7 ] . A key limitation in these scenarios is degeneration, a phenomenon that occurs when point cloud data lacks sufficient geometric variation to constrain full 6-DoF pose estimation. This often arises in feature-scarce or repetitive environments where methods like iterative closest point (ICP) struggle to align consecutive scans [ 8 , 9 , 10 , 11 , 12 , 13 , 14 ] . As a result, optimization becomes ill-conditioned, leading to pose drift and accumulated mapping errors. Detecting and mitigating degeneracy is essential to maintaining SLAM accuracy under such conditions. To address these challenges, recent research has explored several strategies. Some approaches dynamically adjust the contribution of point cloud features based on their reliability, improving alignment in degenerate areas [ 12 ] . Others use probabilistic techniques to assess uncertainty in pose estimation and prevent overconfident updates in weakly constrained directions [ 9 ] . Geometric analyses, such as examining the eigenvalues of the Hessian matrix, have also been applied to identify and respond to degenerate conditions. Additionally, denoising methods and sensor fusion—incorporating cameras or inertial measurement units (IMUs)—have been employed to provide complementary constraints and enhance robustness [ 10 ] . Despite these advances, existing methods often rely on empirically tuned parameters or augmentations to traditional ICP frameworks, which remain sensitive to noise and prone to local minima. And the root causes of degeneration—such as geometric ambiguity and sensor noise—remain underexplored, limiting the generalizability of current solutions. Furthermore, while integrating additional sensing modalities can alleviate the effects of LiDAR degeneracy, the fundamental challenge lies in developing more robust and reliable strategies to directly address LiDAR degeneracy, rather than bypassing it unless absolutely necessary. There is a pressing need for a reimagined registration framework that is both noise-resilient and capable of adapting to geometric degeneracy. This paper presents a comprehensive analysis of LiDAR SLAM degeneration in unstructured environments and proposes a novel LiDAR-IMU front-end matching pipeline to enhance robustness under such conditions, as shown in Fig. 1 . We examine the degeneration problem from three perspectives: • Low observability of state variables : As illustrated in Fig. 3 of [ 23 ] , certain environment types lead to unobservable variables that degrade SLAM performance. • Point cloud mismatches : Unstructured environments tend to cause point cloud mismatches during the ICP process, resulting in errors in the formulation of the optimization • Ill-conditioned optimization : In these scenarios, the ICP cost function tends to exhibit flat minima, making the optimization process highly sensitive and prone to error. The low observability of state variables is fundamentally dictated by environmental geometry and thus lies beyond algorithmic control. To tackle the remaining two challenges, we propose a dynamic, per-point thresholding mechanism for robust correspondence filtering, and an adaptive fusion framework that integrates geometric and inertial constraints to strengthen registration under degenerate conditions. Our contributions are as follows: 1. We introduce a dynamic, per-point distance thresholding mechanism that selectively filters out invalid correspondences by accounting for both the sensor-to-point distance and platform motion, effectively improving matching robustness in degenerate scenarios. 2. We develop an optimization framework that fuses geometric and inertial constraints using an adaptive weighting scheme, enabling robust pose estimation even under weak geometric conditions. 3. We develop an improved LIO front-end and integrate it into a modified LIO-SAM framework for system-level evaluation. By benchmarking the proposed method against existing LiDAR-SLAM systems, we demonstrate its enhanced robustness in feature-scarce environments. II RELATED WORK II-A Robust Point Cloud Registration Point cloud registration is a fundamental component of LiDAR-based SLAM, with the ICP algorithm being the most widely used approach [ 14 ] . ICP iteratively estimates the transformation between point clouds by establishing correspondences—typically via nearest-neighbor searches—and minimizing a cost function that reflects alignment error. To better handle the sparsity and structure of point cloud data, various error metrics such as point-to-point [ 14 ] and point-to-plane [ 15 ] have been employed to improve registration stability and accuracy. However, traditional ICP is sensitive to outliers and noise, which can severely degrade performance. To address this, robust M-estimators—such as Cauchy, and Huber—are incorporated via loss functions to down-weight outlier influence and enhance robustness [ 16 ] . To further tackle ICP’s tendency to get trapped in local minima, globally optimal methods such as TEASER [ 17 ] , based on graduated non-convexity, have gained traction for their ability to handle high outlier ratios and ensure reliable registration. Beyond classical ICP-based methods, alternative registration strategies have been developed to improve robustness in challenging environments. Redundancy-aware approaches such as RMS [ 18 ] reduce point redundancy to enhance real-time pose estimation under geometric degeneracy. Other directions include feature-based methods that extract salient structures [ 19 ] , probabilistic models grounded in statistical inference [ 20 ] , and learning-based techniques that improve adaptability across diverse scenarios [ 21 ] . II-B LiDAR-inertial Odometry LiDAR-inertial odometry (LIO) is central to SLAM and is typically implemented through either optimization-based or filtering-based frameworks. Optimization-based methods—exemplified by LOAM [ 2 ] and its successors such as LIO-SAM [ 3 ] employ scan matching strategies to align sparse geometric features and estimate poses. These