Guest
Submitted to IEEE Transactions on Robotics. X-ICP: Localizability-Aware LiDAR Registration for Robust Localization in Extreme Environments Turcan Tuna, Julian Nubert, Yoshua Nava, Shehryar Khattak, Marco Hutter Paper: Project Website: Abstract: Modern robotic systems are required to operate in challenging environments, which entails reliable localization under various conditions. A taxing scenario for LiDAR-based localization is feature scarcity, which can prompt geometry dependant algorithms like the Iterative Closest Point (ICP) to output wrong estimates and to diverge along weakly constrained directions. To overcome this issue, this work proposes a complete framework that i) does robust multi-category (non-)localizability detection and ii) localizability-aware constrained ICP optimization. The proposed method achieves these feats by utilizing the correspondences between the scan and the map to analyze the alignment strength against the principal directions of the optimization as part of its multi-category LiDAR localizability analysis. This localizability analysis is then tightly integrated into the scan-to-map point cloud registration to allow drift-free pose updates only along well-constrained directions. While state-of-the-art methods often fail to detect rotational non-localizability and struggle with translational non-localizability, the proposed approach can successfully operate in all tested scenarios without environment-specific parameter tuning. The presented approach is put into context and thoroughly compared to past research. The advantages of the presented approach are demonstrated through a series of experiments, underlying the gain in performance and reliability in difficult real-world scenarios.
