SymboSLAM: Semantic Map Generation in a Multi-Agent System

TL;DR

SymboSLAM tackles the problem of explainability in environment-type classification for SLAM by fusing sub-symbolic feature extraction with symbolic ontology reasoning within a multi-agent system. Edge agents perform local SLAM and semantic labeling, while a central control agent merges maps and uses a semantics engine and ontology to produce probabilistic environment-type classifications on a semantically annotated 2D map. The approach is evaluated in simulated EyeSim environments and real-world Canberra-area trials, reporting metrics such as IoU and AP to assess both perception and symbolic reasoning performance. The work demonstrates the feasibility of transparent, ontology-grounded environment interpretation in SLAM, highlights current SLAM limitations under real-world sensing, and outlines concrete paths for improving landmark representations, signage-based context, and hierarchical knowledge chaining to enhance trust and applicability.

Abstract

Sub-symbolic artificial intelligence methods dominate the fields of environment-type classification and Simultaneous Localisation and Mapping. However, a significant area overlooked within these fields is solution transparency for the human-machine interaction space, as the sub-symbolic methods employed for map generation do not account for the explainability of the solutions generated. This paper proposes a novel approach to environment-type classification through Symbolic Simultaneous Localisation and Mapping, SymboSLAM, to bridge the explainability gap. Our method for environment-type classification observes ontological reasoning used to synthesise the context of an environment through the features found within. We achieve explainability within the model by presenting operators with environment-type classifications overlayed by a semantically labelled occupancy map of landmarks and features. We evaluate SymboSLAM with ground-truth maps of the Canberra region, demonstrating method effectiveness. We assessed the system through both simulations and real-world trials.

Figures (11)

• Figure 1: SymboSLAM Edge Agent architecture for a simulated robotic platform. Note that the physical instantiation of an edge agent holds the same basic topology but utilises GPS instead of an IMU.
• Figure 2: Map matching technique employed by the SymboSLAM architecture at both an individual and collective level. Note that the nodes shown within each map are an entry within a pose graph and then displayed to the operator as a point on a 2D occupancy map.
• Figure 3: SymboSLAM ontology. Environment types entries reflect the Canberra region. The SymboSLAM ontology also features concepts from the OntoSLAM and ONTO4MAT ontologies to enable more effective SLAM and swarm control functionality, respectively.
• Figure 4: SymboSLAM symbolic reasoning modules. The ontology featured is the SymboSLAM ontology.
• Figure 5: SymboSLAM control agent architecture.