Multi S-Graphs: An Efficient Distributed Semantic-Relational Collaborative SLAM

TL;DR

The paper tackles scalable, accurate CSLAM in structured environments where exchanging low-level measurements is costly and loop closures are prone to errors. It introduces Multi S-Graphs, a distributed CSLAM framework built on four-layered, optimizable S-Graphs that blend pose graphs with a hierarchical 3D scene graph and semantic-relational constraints. A Room Descriptor that fuses semantic room information with Room Centric point clouds enables robust inter-robot loop closures, while S-Graph Distillation minimizes data exchange. Experiments in simulation and real-world settings demonstrate significant improvements in accuracy and substantial bandwidth reductions compared with state-of-the-art baselines, highlighting the approach’s potential for scalable multi-robot mapping in structured environments.

Abstract

Collaborative Simultaneous Localization and Mapping (CSLAM) is critical to enable multiple robots to operate in complex environments. Most CSLAM techniques rely on raw sensor measurement or low-level features such as keyframe descriptors, which can lead to wrong loop closures due to the lack of deep understanding of the environment. Moreover, the exchange of these measurements and low-level features among the robots requires the transmission of a significant amount of data, which limits the scalability of the system. To overcome these limitations, we present Multi S-Graphs, a decentralized CSLAM system that utilizes high-level semantic-relational information embedded in the four-layered hierarchical and optimizable situational graphs for cooperative map generation and localization in structured environments while minimizing the information exchanged between the robots. To support this, we present a novel room-based descriptor which, along with its connected walls, is used to perform inter-robot loop closures, addressing the challenges of multi-robot kidnapped problem initialization. Multiple experiments in simulated and real environments validate the improvement in accuracy and robustness of the proposed approach while reducing the amount of data exchanged between robots compared to other state-of-the-art approaches. Software available within a docker image: https://github.com/snt-arg/multi_s_graphs_docker

Figures (6)

• Figure 1: Collaborative graph generated by three robots mapping a building floor (orange, gray, and blue for robots 1, 2, and 3 respectively). The collaborative graph is generated from the point of view of Robot 1. Semantic information gathered by the other robots, such as rooms or walls appears in transparent colors in the graph layers. The origin of the coordinate systems of the other robots appears surrounded by a circle. Below is the result of merging the point clouds obtained from each of the robots. This point cloud is displayed for visualization purposes and is not utilized in inter-robot communication.
• Figure 2: Multi S-Graph architecture viewed from $Robot_{i}$ perspective. Odometry measures and 3D Lidar data are the inputs that the robotic platform provides. Each robot executes its own Local S-Graph system and generates the Room Descriptors and the Distilled S-Graphs that are exchanged with the rest of the robots. Messages exchanged between robots are represented in gray. From the point of view of the $Robot_{i}$ the rest of the agents provide the Distilled graphs and the Room descriptors that are used to generate the Collaborative S-Graph. Information provided by the rest of the robots appears in green color.
• Figure 3: Room Descriptor (right) obtained from a Room Centric point cloud (left) by using a Scan Context. The value of each cell in the descriptor is the maximum height of the points within a corresponding bin.
• Figure 4: Simulated experiments, showing the trajectory and the point cloud map generated for each robot.
• Figure 5: Real experiment scenarios, showing the trajectory and the point cloud map generated for each robot. In (e) the experiment was run with 3 robots.