A Methodological Framework for Runtime Ontology Evolution in Dynamic Environments

Intelligent systems operating in real-world environments are often required to make decisions in contexts that are only partially known at design time. In such scenarios, the assumption of a static and fully specified knowledge base becomes unrealistic, limiting the system’s ability to adapt to novel situations. This challenge is particularly relevant for robotic systems, whose behavior cannot be entirely pre-programmed when operating in dynamic and evolving environments. This paper proposes a methodological and architectural approach for the runtime update of ontologies and knowledge bases, enabling intelligent systems to autonomously adapt their internal representation of the world during execution. The proposed approach enables the system to identify knowledge gaps by distinguishing between previously unknown concepts and known concepts enriched with newly observed instances, and to integrate such information into the ontology in a controlled and consistent manner. The approach is implemented as an end-to-end pipeline that combines visual perception, semantic interpretation through large language models, and a robust ontology update mechanism. Particular attention is devoted to ensuring formal consistency during runtime evolution, addressing challenges such as the generation of valid OWL constructs, the management of inverse properties, datatype normalization, and the prevention of semantic degradation over iterative updates. By enabling knowledge-driven adaptation at runtime, the proposed framework supports autonomous decision-making in environments that cannot be fully anticipated at design time. The approach was developed within the MUSIC4D and MHARA projects, which explore the use of intelligent systems in dynamic, partially structured contexts, focusing on knowledge-based adaptation.