A Deep Learning Approach for 2D Lidar-based localization of an autonomous Quanser QCar

Accurate and reliable localization is a fundamental requirement for autonomous mobile robots operating in structured environments. Traditional approaches, such as Monte Carlo Localization (MCL), rely on probabilistic filtering to fuse odometry and LiDAR measurements with a prior map, achieving robust performance at the cost of hand-engineered models and assumptions. In this work, an alternative data-driven solution is proposed, where localization is addressed through a neural network architecture designed to directly estimate the robot pose within a known occupancy grid map. The central idea is to fuse heterogeneous information from odometry and LiDAR scans with the spatial structure of the map using a cross-attention mechanism, enabling the network to jointly exploit temporal and spatial cues. The design explores the combination of convolutional layers for map encoding, sequence models for sensor data, and attention modules for modality integration. A dedicated dataset was collected using a motion capture system to provide high-precision ground truth, ensuring reliable supervision during training. The objective of this study is to investigate whether neural networks can represent a viable alternative to classical probabilistic methods for robot localization, focusing on architectural design, computational feasibility, and the integration of heterogeneous sensory inputs.