Multi S-Graphs: an Efficient Real-time Distributed Semantic-Relational Collaborative SLAM

Collaborative Simultaneous Localization and Mapping (CSLAM) is critical toenable multiple robots to operate in complex environments. Most CSLAMtechniques rely on raw sensor measurement or low-level features such askeyframe descriptors, which can lead to wrong loop closures due to the lack ofdeep understanding of the environment. Moreover, the exchange of thesemeasurements and low-level features among the robots requires the transmissionof a significant amount of data, which limits the scalability of the system. Toovercome these limitations, we present Multi S-Graphs, a decentralized CSLAMsystem that utilizes high-level semantic-relational information embedded in thefour-layered hierarchical and optimizable situational graphs for cooperativemap generation and localization while minimizing the information exchangedbetween the robots. To support this, we present a novel room-based descriptorwhich, along with its connected walls, is used to perform inter-robot loopclosures, addressing the challenges of multi-robot kidnapped probleminitialization. Multiple experiments in simulated and real environmentsvalidate the improvement in accuracy and robustness of the proposed approachwhile reducing the amount of data exchanged between robots compared to otherstate-of-the-art approaches. Software available within a docker image:https://github.com/snt-arg/multi_s_graphs_docker