Real-Time Defect Detection, Analysis and Suggestion in Adaptive 3D Concrete Printing via Multimodal LLM Integration (2025-10)¶

3D Concrete Printing faces challenges such as low adaptability to unpredictable changes in the environment, which potentially affects not only the 3D-printed concrete quality but also the structural performance of the construct. This paper proposes integrating Multimodal Large Language Models (mLLMs) into 3D Concrete Printing systems for real-time quality detection. The framework employs mLLM to analyse synchronised visual feeds and sensor data (e.g., extrusion rates and layer adhesion metrics) to identify deviations such as appearance, surface inaccuracies, and geometric deformations. The language model synthesises inputs to generate text-based commands for process parameter adjustments (e.g., materials formulation, nozzle position and speed, material flow). mLLM-assisted horizontal shifting detection simulations are conducted, demonstrating the feasibility through prompt engineering. It is foreseen that coupling sensor fusion with contextual language understanding can enable autonomous error mitigation, reduce manual intervention, enhance process reliability, and create the conditions for standardisation and industrialisation of 3DCP. This integration suggests mLLMs’ potential to advance adaptive, self-correcting 3DCP systems, supporting efficient and automated construction practices.