Towards an AI-Based Framework for Autonomous Design and Construction: Learning from Reinforcement Learning Success in RTS Games

The present study summarizes the state-of-the-art research in deep reinforcement learning (DRL) techniques in the architecture, engineering and construction industry and it formulates a general framework for autonomous design and construction. The framework is inspired by the noticeable success of DRL and imitation learning algorithms in real time strategy (RTS) games, which normally require efficient resource planning and long-term strategic coordination. The objective of the proposed framework is to reduce data segregation and loss of project information. The prevention of data leakage is achieved by replacing the linear process with an iterative one where the consequences of design decisions on the construction process (and vice versa) are understood in a virtual environment simultaneously. The proposed framework also exploits recent advances in simulated physics-based environments, like game engines. Designers and builders can therefore simulate on-site scenarios and exchange views on the required design and construction goals early in the project. The multi-objective optimization problem is then passed to artificial agents. These agents train on achieving the project goals under the supervision of a team of humans. The tacit knowledge transferred to the brain of the agents can later be deployed on-site through execution robots. The proposed approach is demonstrated by a proof-of-concept software application, showcasing a brick pavilion project. Design and construction constraints are first imposed by the user. Agents are then trained using a DRL algorithm.