PASTEL: An Aerial Multi-LiDAR Dataset for Research in SLAM Tuning and Robustness

Abstract

LiDAR-based SLAM algorithms are critically dependent on the configuration of the environment and the LiDAR characteristics. Existing LiDAR-based datasets are not devised for research in SLAM parameter tuning due to limited diversity in types of environments and difficulties in comparing results from LiDARs with different characteristics. This paper presents PASTEL, a dataset for research in LiDAR-based SLAM tuning performance and robustness for low-altitude navigation of uncrewed aerial vehicles. PASTEL provides high diversity in the main factors that affect LiDAR-based SLAM tuning: type of environment, 3D LiDAR characteristics, and aerial robot velocity. The aerial robot mounts three different 3D LiDAR models (with different resolutions, fields of view, and accuracies) that are recorded in parallel. PASTEL includes 17 sequences recorded in flights with two velocity profiles at distinct types of environments: wide-open spaces with distant obstacles, horizontally confined spaces with open sky, confined GNSS-denied spaces, and also includes sequences where the aerial robot transitions between two or more of these distinct environments. The dataset includes measurements from the 3 LiDARS and Inertial Measurement Units (IMUs), as well as the ground-truth robot trajectories and ground-truth maps. PASTEL, available at https://sites.google.com/view/pastel-lidar-slam-dataset and at https://zenodo.org/records/14796964, has been validated in terms of trajectory and map accuracies with well-known SLAM methods.