Towards autonomous shotcrete construction: semantic 3D reconstruction for concrete deposition using stereo vision and deep learning

The adoption of autonomous systems is a foreseeable necessity in the construction sector due to work hazards and labor shortages. This paper presents a semantic 3D understanding module that creates 3D models of construction sites with highlighted regions of interest for shotcrete application. The approach uses YOLOv8m-seg and SiamMask for robust semantic segmentation together with RTAB-Map and InfiniTAM for visual odometry and 3D reconstruction. Our method is the first step towards a novel, autonomous robot for shotcrete spraying and finishing. The effectiveness of our approach is shown on a mock-up construction site and provides evidence for the applicability of robotic construction