Distributed Lidar-based Simultaneous Localization and Mapping

Simultaneous Localization and Mapping (SLAM) algorithms provide a robust solution for mobile robots localization and map building of surrounding environment even when most used positioning systems (GPS) are not available for autonomous navigation, such as in indoor or subterranean locations. Being able to set a team of robots to resolve this common task and enable it to collaborate, it is possible to obtain better results in shorter time. In this thesis work, carried out in collaboration with LINKS Foundation, a fully distributed collaborative SLAM system, based on lidar sensor data processing, has been designed and partially developed exploiting the ROS2 open-source framework. A distributed approach, where every robot is implied on perception and optimization tasks, is selected according to robustness, scalability and security requirements, facing with algorithm complexity. The solution is modeled in five blocks following the typical structure characterising implemented systems in recent research. The point clouds coming from sensors are firstly processed by the Signal Processing module to provide odometry and a downsampled representation of data, useful to ease successive elaborations. The odometry gives ego-motion of a robot, here lidar-based algorithms are chosen in order to exploit the better results regarding this first localization information as compared to other sensor suites. In order to reduce estimation error caused by odometry drift, robots must recognize whether a position is already crossed, detecting the so-called Loop Closures. A robot within a team can perform the detection during its navigation (Intra-Robot Loop Closure detection) or during a rendezvous (Inter-Robot Loop Closure detection). In the latter case the robot shares its estimations and compares visited locations with those of the neighbor. Local and collaborative versions of modules are developed using Scan Contexts: point cloud descriptors, compact and sufficiently representative, used to save computational and communication resources. Errors on detection can occur due to perceptual aliasing: the set of loops is filtered by the Outlier Rejection module. Finally the resultant loops, together with odometry information, are processed by the Distributed Mapper which optimizes trajectories and updates the maps of the entire team, reaching consensus on final estimation.