Abstract
SlideSLAM is a real-time, decentralized SLAM framework designed for heterogeneous robot teams, such as drones and ground rovers, to collaboratively build sparse, object-based maps. Unlike traditional dense geometric reconstructions, SlideSLAM represents the environment using geometric primitives (cuboids, cylinders, ellipsoids), significantly reducing communication bandwidth and enabling robust map merging across different sensor modalities. This paper details the modular architecture, including semantic front-ends and a decentralized back-end, and evaluates the system's performance in diverse real-world environments.
Introduction
Simultaneous Localization and Mapping (SLAM) is fundamental for autonomous navigation. However, traditional methods often produce dense point clouds or voxels that are computationally expensive to share and lack semantic meaning for high-level tasks. SlideSLAM addresses three critical gaps: heterogeneity in sensor data (e.g., LiDAR vs. RGBD), communication constraints in real-world deployments, and the risks associated with centralized systems. By utilizing semantic objects as a universal language, SlideSLAM enables robust, decentralized collaboration.
Methodology
The SlideSLAM architecture consists of a modular front-end for perception and a decentralized back-end for optimization.
Perception and Geometric Modeling
The system supports LiDAR (RangeNet++) and RGBD cameras (YOLOv8/YOLO-World). Detected objects are abstracted into simple geometric shapes: cuboids for vehicles, cylinders for trees or poles, and ellipsoids for irregular objects. This abstraction ensures viewpoint and sensor invariance.
Decentralized Optimization and Loop Closure
A factor-graph-based approach using iSAM2 optimizes robot trajectories and object landmarks. For place recognition, SlideSLAM introduces two algorithms:
- SlideMatch: An exhaustive search for loop closures.
- SlideGraph: A descriptor-based approach using Delaunay triangulation to create unique signatures of object clusters, significantly reducing the search space for loop closures.
Experimental Evaluation
SlideSLAM was evaluated in four distinct environments: the Pennovation Campus (outdoor urban), indoor lab buildings (cluttered aerial-ground collaboration), forest environments (high-density landmarks), and the Semantic KITTI benchmark.
Results and Discussion
SlideSLAM demonstrates exceptional communication efficiency, requiring less than 10 KB for maps that typically consume megabytes. In urban settings, inter-robot localization achieved a position error of ~0.13m. The system also showed high reliability in object mapping, with F1 scores exceeding 0.92 for cars and tree trunks.
The sparsity of the map is a key feature, providing actionable data for navigation while maintaining a minimal data footprint. However, the current reliance on three basic shapes and the assumption of a static world present opportunities for future improvement, such as integrating Large Language Models (LLMs) to reason about dynamic objects.
Conclusion
SlideSLAM provides a scalable, decentralized solution for multi-robot semantic SLAM. By bridging the gap between raw sensor data and high-level semantic understanding through sparse geometric primitives, it enables efficient and robust collaborative navigation in complex environments.
Summary updated