A comparative evaluation of LiDAR Odometry Algorithms in Urban and Unstructured scenarios

LiDAR-based odometry is a cornerstone of autonomous navigation, robotic perception, and mapping, offering precise pose estimation even when cameras or GPS become unreliable. However, most existing algorithms are designed and benchmarked for structured, urban-like environments, leaving their robustness in unstructured settings—such as vineyards, forests, and off-road terrains—insufficiently validated. This gap poses a critical limitation for deploying autonomous systems in agricultural robotics and other non-urban applications where localization failures can compromise safety and mission success. Four state-of-the-art LiDAR odometry algorithms—KISS-ICP, Genz-ICP, MOLA-LO, and SiMpLe—were evaluated on two contrasting datasets. The KITTI benchmark represented structured urban roads, while a custom vineyard dataset collected by the Interdepartmental Center PiC4SER captured the irregular and repetitive geometry of agricultural fields. Each algorithm was first tested using its default configurations and then re-optimized for both environments. Performance was assessed through standard KITTI odometry metrics (absolute and relative pose error) and qualitative trajectory comparisons against GPS ground truth. On the KITTI sequences, all methods demonstrated competitive accuracy, with KISS-ICP and MOLA-LO exhibiting slightly lower drift. In the vineyard dataset, performance differences became pronounced: Genz-ICP, which relies heavily on planar and edge features, experienced substantial accuracy degradation; SiMpLe achieved high computational efficiency but suffered frequent misalignments in repetitive row patterns; and KISS-ICP maintained stable estimates through its adaptive correspondence strategy. The findings reveal that benchmarking exclusively on urban data overestimates odometry robustness in unstructured environments. Hybrid strategies combining feature-rich techniques with adaptive correspondence models—such as those employed by MOLA-LO—show strong potential for enhancing reliability in agricultural robotics and similar contexts. Expanding benchmarking datasets to include diverse, real-world scenarios is essential for advancing LiDAR odometry toward dependable deployment beyond controlled urban settings.