SGM-SLAM is a distributed SLAM framework for robot teams with LiDAR, cameras and IMUs that finds inter-robot constraints by matching scene graphs. Each graph holds a semantically segmented point-cloud layer and a layer of bounded objects, and matching uses only object labels and centroids, which keeps the exchanged data small. A multi-step exchange and optimisation scheme cuts communication further; it is tested in simulation and on legged robots indoors and outdoors.
Relevance
Matching on labels and centroids is cheap, and it also means one robot’s wrong label can become a wrong loop closure; the abstract does not say how mismatches are rejected. Reading for collaborative SLAM and 3D scene graphs in multi-robot perception.