Autonomous modular construction strategy using robotized crane based on deep learning and reinforcement learning

Abstract

Modular construction offers significant advantages including faster construction time, higher quality control and less environmental impact. To further enhance its advantages, advanced robotic construction technologies are being developed. This research develops an automated modular construction framework that incorporates the robotic kinematics, deep learning and deep reinforcement learning using a robotized crane. The proposed modular construction strategy utilizes YOLOv5-S for modular container identification and localization. An improved proximal policy optimization (PPO-I) is developed and implemented in this strategy for collision-free three-dimensional (3D) lifting path planning and modular container transportation. States and rewards of the PPO-I and robot kinematics design of a real mobile crane are developed. The feasibility of the proposed modular construction strategy is verified through four case studies in 3D virtual environments. More than 97% success rate is observed meaning that the proposed strategy can be implemented in the robotized crane to localize the modular container and transport it to the target position with collision avoidance. The results indicate the potential of the proposed robotic-assisted modular construction strategy in the field of automated construction.