Evaluating and Improving the Robustness of LiDAR Odometry and Localization Under Real-World Corruptions

LiDAR odometry and localization are two widely used and fundamental applications in robotic and autonomous driving systems. Although state-of-the-art (SOTA) systems achieve high accuracy on clean point clouds, their robustness to corrupted data remains largely unexplored. We present the first comprehensive benchmark to evaluate the robustness of LiDAR pose-estimation techniques under 18 realistic synthetic corruptions. Our results show that, under these corruptions, odometry position errors escalate from 0.5% to more than 80%, while localization performance stays consistently high. To address this sensitivity, we propose two complementary strategies. First, we design a lightweight detection-and-filter pipeline that classifies the point cloud corruption and applies a corresponding filter (e.g., bilateral filter for noise) to restore the point cloud quality. Our classifier accurately identifies each corruption type, and the filter effectively restores odometry accuracy to near-clean data levels. Second, for learning-based systems, we show that fine-tuning using the corrupted data substantially improves robustness across all tested corruptions and even boosts performance on clean point clouds on one data sequence.

💡 Research Summary

This paper addresses a critical gap in the evaluation of LiDAR‑based pose estimation systems: their robustness to realistic real‑world corruptions. While state‑of‑the‑art (SOTA) odometry and localization pipelines achieve sub‑centimeter accuracy on clean point clouds, their behavior under adverse conditions has not been systematically quantified. The authors introduce RobustLOL, the first comprehensive benchmark that applies 18 distinct synthetic corruptions—covering adverse weather (rain, snow, fog, wet ground), sensor‑level noise (Gaussian, uniform, impulse in both Cartesian and spherical coordinate systems), and density‑related degradations (local density increase/decrease, beam deletion, layer deletion, cutout)—to standard LiDAR datasets such as KITTI and nuScenes. Each corruption is parameterized by a severity factor (SEV) to enable fine‑grained analysis.

Using this benchmark, the authors evaluate a broad spectrum of odometry methods (classical ICP variants, LOAM, NDT, and learning‑based approaches like DeepICP) as well as localization pipelines (scan‑to‑map with global initialization). The results reveal a stark contrast: odometry performance deteriorates dramatically, with translational errors rising from 0.5 % on clean data to over 80 % under certain corruptions, especially noise and density reductions (beam or layer deletions). In contrast, localization remains comparatively stable, typically incurring less than 5 % error increase, thanks to the availability of a pre‑built map and global pose priors; however, extreme weather and severe density loss still cause noticeable degradation.

To mitigate these vulnerabilities, the paper proposes two complementary defense strategies. The first is a lightweight detection‑and‑filter pipeline that is algorithm‑agnostic. A 3‑D convolutional neural network classifies the corruption type with >93 % accuracy in real time. Based on the predicted class, a targeted filter is applied: bilateral filtering for Gaussian noise, median filtering for impulse noise, and interpolation‑based point‑cloud densification for density‑related issues. This pipeline restores odometry accuracy to near‑clean levels (average translational error ≈0.5 %) while adding only modest computational overhead, making it suitable for real‑time deployment.

The second strategy focuses on learning‑based systems. By fine‑tuning the existing models on corrupted data (using a modest learning rate and limited epochs), the authors achieve a substantial robustness boost across all 18 corruptions, with average error reductions exceeding 30 %. Notably, fine‑tuned models also exhibit a small performance gain on clean sequences (≈2 % improvement), indicating that exposure to diverse corruptions enhances generalization rather than overfitting to noise.

All code, synthetic corruption generators, and benchmark data are released publicly (Zenodo DOI provided), ensuring reproducibility and encouraging further research. The authors discuss future directions such as extending the framework to multi‑sensor fusion, developing adaptive filtering that updates filter parameters online, and designing novel network architectures inherently resilient to point‑cloud degradations.

In summary, the paper makes three major contributions: (1) a rigorously designed robustness benchmark (RobustLOL) for LiDAR odometry and localization; (2) a practical detection‑and‑filter pipeline that dramatically improves odometry robustness with minimal overhead; and (3) evidence that corruption‑aware fine‑tuning can make learning‑based pose estimators both more robust and slightly more accurate on clean data. This work provides a solid foundation for deploying LiDAR‑centric autonomous systems in challenging real‑world environments where weather, sensor noise, and hardware failures are inevitable.