Introduction. Robotics applications require robust mechanisms to represent and reason about knowledge in dynamic environments. Prolog has long been considered an efective symbolic language for robotics due to its declarative nature and support for reasoning. However, creating structured KBs manually is challenging. Recent advances in LLMsprovide new opportunities for extracting structured knowledge from unstructured input. Our research investigates how Prolog and LLMs can be combined into a unified framework for multi-agent plan generation and execution with Behavior Trees (BT). Framework Overview. The proposed architecture follows three main steps: (1) Knowledge Base Generation with LLM Support . The KB encodes facts and actions in Prolog. States are described using predicates such as on(b1,b2,pos1) or av(ag1). Actions are defined through preconditions and efects. LLMs are employed to translate natural language queries into Prolog clauses describing initial and goal states (see Fig. 2 in the full paper).
CEUR