Feature-Level Fusion of RGB Camera and 2-D LiDAR for Local Occupancy Estimation and Reactive Navigation: A Controlled Simulation Study

Abstract

This preprint presents a controlled, systems-level simulation study of feature-level fusion of RGB camera and 2-D LiDAR for local occupancy estimation and reactive navigation of an indoor mobile robot. The system integrates ROS 2, PyTorch and Gazebo in a modular perception-to-navigation pipeline. Three perception configurations are considered: camera-only, LiDAR-only and feature-level RGB-LiDAR fusion. The configurations use a common 20×20 local occupancy representation, shared dataset split, downstream interface, navigation controller and evaluation protocol. The fusion model combines a 512-dimensional camera representation with a 256-dimensional LiDAR representation through a learned feature-level mapping. The study reports quantitative perception, localisation, inference-latency and closed-loop navigation results, together with a separately evaluated short-term prediction extension. The work is intentionally scoped to a bounded indoor Gazebo simulation and does not claim a novel state-of-the-art fusion algorithm, real-world deployment, sim-to-real validation, SLAM, reinforcement-learning-based navigation or full three-dimensional scene understanding. The accompanying software archive provides the project implementation and supplementary reproducibility materials associated with the manuscript.