Localização e mapeamento para robôs móveis em ambientes confinados baseado em fusão de LiDAR com odometrias de rodas e sensor inercial

Abstract

The exploration and inspection of confined environments using robotic devices is a feasible and safe option. However, confined environments present several challenges to robotics, especially regarding pose estimation and map generation which are essential for autonomous navigation. Due to the lack of GPS signal and possible interference with magnetometers, especially in the mining region or in the presence of high-powered electrical equipment, it is impossible to use these sensors to estimate the position and orientation of the robotic device. Low lighting can impair the identification of visual features, and slippery terrains lead to the accumulation of wheel odometry errors. Thus, a more suitable solution is the Simultaneous Localization And Mapping based on LiDAR sensors, called LiDAR SLAM. For the proper functioning of an autonomous robot, the location needs to be estimated online in order to feedback the navigation control system, and the map must be representative and computationally lighweight, facilitating the execution of the path planning embedded algorithms. Therefore, this dissertation presents a study and investigation of 3 state-of-the-art techniques, LOAM-Velodyne, LeGO-LOAM, and HDL-Graph-SLAM, to determine which one is the most suitable for implementation in the EspeleoRobô, a robotic device for inspecting confined environments. This mobile robot is being developed by the InstitutoTecnológico Vale (ITV) in partnership with the Universidade Federal de Minas Gerais (UFMG). The comparison between LiDAR SLAM techniques uses metrics proposed to assess the accuracy of localization estimations and maps generated during simulations in virtual environments implemented with the CoppeliaSim software, and also in real-world experiments. The best performing technique, LeGO-LOAM, was adapted and embedded into the EspeleoRobô, allowing to evaluate the online performance of localization and mapping during indoor and outdoor tests at UFMG and in field experiments at Mina duVeloso. Given that some of the environments explored by EspeleoRobô, including ducts and galleries, present few geometric features that impact the performance of LiDAR SLAM techniques, generating an underestimation of the robot traveled distance and consequently the deformation of the generated map, we propose an Extended Kalman Filter (EKF) integrated to the SLAM technique. This filter merges data from the LiDAR odometry, which uses an adaptive covariance in respect to the number of environment features identify, the wheel odometry and the IMU available in the robot. The implementation was initially evaluated with simulations in virtual environments, and then validated during real-world experiments with the EspeleoRobô.