Open Collections

Abstract

The construction industry is serving as a cornerstone of economic growth. Despite its significance, the industry faces significant challenges, including a growing demand for labor productivity and workspace safety, particularly in the wake of a pronounced decline in labor productivity following the pandemic. These challenges underscore the urgent need for innovative solutions, with robotics emerging as a technology poised to revolutionize current construction practices. Robots have demonstrated their potential to enhance labor productivity across industries and are expected to perform heavy and repetitive tasks traditionally undertaken by human workers in construction. However, unlike controlled factory environments, the dynamic, sequential, and time-varying nature of on-site construction activities presents challenges to robotic learning. Moreover, the interaction between robots and heavy structure components made by steel, concrete, and wood, combined with the use of tools like drills and saws, introduces new safety risks for workers sharing the workspace with robots. To address these challenges, this study proposes a biomimetic deep reinforcement learning-based control framework designed to enhance learning efficiency and safety for on-site construction using robots. The framework comprises three layers. The low-level pre-stabilized layer generates robot motions in response to control signals from the higher layers. The central nervous system, forming the higher control layers, includes a neural network layer "brain" and a safety-critical layer "spinal cord." The brain receives sensor observations and then generates high-level control signals to guide the robot in accomplishing complex construction tasks. The spinal cord acts as a high-frequency safety layer, ensuring collision-free human-robot collaboration. To enhance learning efficiency, an intuitive demonstration collection platform is introduced for neural network policy learning. Additionally, an imitation learning approach that integrates environmental and intrinsic rewards is proposed to guide the learning process more effectively. Finally, to facilitate collaboration between construction robots, a multi-agent deep reinforcement learning method is proposed. The proposed framework, including the safety-critical layer, imitation learning method, and multi-agent reinforcement learning method, as well as the demonstration collection platform, has been validated through extensive simulation tests and real-world robot experiments. Results highlight the potential of biomimetic control frameworks to advance productivity and safety in the construction industry.