PLANTOR: LLM-Driven Multi-Robot Task Planning with Explainable Knowledge Bases

A new framework named PLANTOR has been developed by researchers to utilize large language models (LLMs) for creating and implementing multi-robot task plans based on natural language inputs. This system builds a structured Prolog knowledge base, performs consistency checks to identify and correct errors, and generates high-level symbolic plans, which are then refined into specific robot actions and optimized schedules that are transformed into executable behavior trees. PLANTOR prioritizes transparency by revealing the knowledge base, intermediate plans, and scheduling constraints. It was tested on scenarios derived from the Blocks World and Grippers benchmarks using various language models, demonstrating both the quality of the knowledge base and runtime efficiency. The framework's end-to-end execution was showcased in a real multi-arm assembly context. The research paper is accessible on arXiv (2502.19135) and has been marked as a revised version. This study tackles the integration of LLM adaptability with symbolic reasoning for effective robotic planning.