LiDAR-Based Deep Learning-Enabled Geometric Fingerprinting for Indoor Robot Localization

This dataset contains all experimental data used in the manuscript “LiDAR-Based Deep Learning-Enabled Geometric Fingerprinting for Indoor Robot Localization” (https://www.mdpi.com/4013946 , Applied Sciences, 2026). The research hypothesis is that handcrafted geometric features extracted from 2D LiDAR scans, combined with a lightweight ConvMLP deep learning model and further smoothened with recursive EKF, can provide accurate and real‑time robot localization in indoor environments without requiring maps or GPU resources. If you use this code, dataset, or methodology in your research, please cite the following peer‑reviewed publication: Keladi Ganapathi, H.; Mukhopadhyay, S. LiDAR-Based Deep Learning-Enabled Geometric Fingerprinting for Indoor Robot Localization. Appl. Sci. 2026, 16, 7718. https://doi.org/10.3390/app16157718 The dataset includes: Three real‑world indoor maps (Map 1, Map 2, and Map 3) in CSV format, containing the robot’s reference pose, corresponding LiDAR scan data, and 11 handcrafted geometric features. ROS2 bag files for all three maps, containing raw LiDAR scans, odometry, TF data, timestamps, and Cartographer SLAM‑based reference poses. Inference ROS2 bag files for all three maps, containing pose predictions generated by the trained ConvMLP model. Preprocessed datasets for each map, including cleaned LiDAR scans, extracted geometric features, augmented samples, and normalized training data. Three trained ConvMLP model files, representing the final learned models used for inference and evaluation. All files are organized by map and data type to enable easy reproduction of the experiments. Please refer to the readme_file document for the file workflow.