Implementation of ROS-based Multi-Agent SLAM Centralized and Decentralized Approaches

With the advent of mobile robotics, new research fields were born. In particular, one of the most interesting challenges regards the so-called Simultaneous Localization and Mapping, also known as SLAM: it concerns the task that a mobile robot accomplishes to accurately create a map of an unknown environment and localize itself in it. Since mapping large environments may take too much time if only a single agent is employed, SLAM frameworks that rely on a swarm of mobile robots begin to be developed in order to not only shorten times for mapping but also achieve higher accuracy. Thus, such methods aim at solving a multi-agent SLAM problem. There are two main ways to map the environment and track the entire squad of agents: either having a central server that does the whole computation or implementing a distributed approach that requires exchange of data between mobile robots. This work presents two different approaches that are inspired by the previous two alternatives and are developed with the help of ROS2, i.e. an open-source set of software tools for robotic applications. Moreover, the mapping is limited to be two-dimensional, hence these methods are specially designed to work with 2D LiDAR sensors. Both frameworks are tested in a real-time scenario where two copies of TurtleBot3 Burger aim at mapping the same indoor environment. The first algorithm is a centralized setting where each agent independently creates its own local map without sending and/or receiving additional data from others: simply, Cartographer locally runs as single-agent SLAM algorithm for each mobile robot and only processes the laser scans of that agent. A central server takes and merges the local maps through feature matching, thus building a global map of the explored environment. Although this naive approach still produces a good global map, the server needs some assumptions on the initial poses of the mobile robots, otherwise the relative pose between the local maps cannot be computed correctly and the feature matching may fail. The second approach is designed as a low-drift, fully-distributed method that incorporates two modules. The single-robot front-end module supports any LiDAR-based SLAM with inertial odometry and performs local intra-robot loop closure. The collaborative aspect of the algorithm lies in the distributed back-end module: during a meeting between two mobile robots, they exchange scan data through LiDAR-Iris, i.e. a robust LiDAR descriptor that lightens the load on the communication link and formats the point cloud into a quickly processable piece of data on which the distributed loop closure recognition operates in order to find possible candidates for inter-robot loop closures. Furthermore, the distributed back-end module performs outlier rejection and an optimization step on the graph that contains the robots' poses in order to refine the produced map.