AI BIM coordinator for non-expert interaction in building design using LLM-driven multi-agent systems

Abstract

The building construction domain relies heavily on human experts for specialized tasks, often leading to excessive workload and inefficiency. For example, building design coordination typically requires BIM coordinators to operate BIM tools on participants' behalf, with limited ability to capture procedural knowledge. To address this gap, this paper introduces AI-based BIM coordinators to enable non-BIM experts to seamlessly interact with BIM tools while automatically documenting valuable procedural information. The approach customizes five LLM-based agents equipped with fundamental building knowledge and skills. Through user prompts, the collaboration of “bim_assistant” and “checker” agents ensures the accuracy and consistency of generated code, while the “sender” and “executor” agents facilitate direct execution in BIM tools beyond text-based responses, with the “terminator” ending the process. An AutoGen-Revit prototype demonstrates feasibility, flexibility, robustness, and time-cost effectiveness for common BIM-based design coordination tasks, including space reasoning, 3D element creation, clash detection, informative 3D view saving, documentation, and retrieval. This paper enables specialist-free workflows and generates procedural logs and images that serve as datasets for future analysis.