Ontology-Guided Multi-Modal Perception for Trusted and Explainable Robotic Action

This paper presents an ontology-guided neuro-symbolic framework that integrates deep visual perception with formal reasoning for safe and explainable robotic behavior. The system maps segmentation outputs from DeepLabV3–ResNet50 into OWL-Lite knowledge graphs validated via SHACL and SPARQL rules. It enables real-time detection of spatial and behavioral violations, generates factual and contrastive explanations, and evaluates compliance using the GOOSE robotics dataset. Experimental results demonstrate over 95 % ontology conformity, a 15 % improvement in safety clearance, and reasoning latency below 10 ms per query.