Comparativa de SLAM RGB-D para manipulación robótica en entornos reducidos

Abstract

This paper experimentally compares three SLAM systems —ORB-SLAM3, RTAB-Map and Cartographer— on RGB-D data in a short-range robotic manipulation setting. Sequences were captured with two Intel RealSense cameras, the D405 and the D435i, mounted on a UR3e robotic arm, which enables repeatable trajectories and reduces the variability typical of hand-held captures. Five trajectory patterns and three scenes of increasing geometric complexity were designed. Processing was performed offline, homogenising the outputs to allow a common comparison in terms of estimated trajectory, 3D map, 2D map, visual interpretability and computational cost. Results show no universally superior system: ORB-SLAM3 with the D405 offers the best compromise for close scenes, RTAB-Map exhibits the highest robustness across sensors, and Cartographer stands out for 2D occupancy mapping. Inertial information yields situational improvements but no general benefit.