Abstract
Localization is one of the fundamental problems in mobile robotics, being the basis of autonomous navigation and trajectory planning. While technologies such as GPS, trilateration, or triangulation are widely used outdoors, they are not applicable in indoor environments. The most common methods to address this problem are SLAM algorithms based on visual sensors, such as RGBD cameras or laser ranging sensors. However, these methods are susceptible to failures in environments with poor illumination and regions without texture or transparent walls. To avoid those issues, a novel alternative is to utilize machine learning methods that take inertial data to compute the position of a mobile robot as it moves in the environment. This work compares seven different deep neural network models to estimate the robot's displacements in the environment and reconstruct its current position: ResNet-18, ResNet-50, RestNet-Small, MobileNetV3-Small, MobileNetV3-Large, CNN + LSTM, and RBCN-Net. We perform experiments in the empty space and two different environments with obstacles. The best results for the empty space were obtained using the MobileNetV3 architecture, with a mean RMS error of 0.9794 through ten test trajectories. ResNet-18 obtained the best results for environments with obstacles, having a mean RMS error of 0.4571.
DOI: 10.61416/ceai.v26i3.9096