Reinforcement learning (RL) addresses sequential, stochastic decision processes across construction engineering and management. This study aims to analyze the intellectual structure, thematic evolution and empirical maturity of RL across the asset lifecycle to inform evidence-based industry adoption.
A corpus of 1,198 peer-reviewed articles from Scopus and Web of Science was systematically screened and analyzed. Scientometric mapping with Louvain community detection delineated thematic clusters, while a three-level rubric classified deployment maturity across simulation, controlled testing and operational field demonstration.
Six research fronts structure the discipline, comprising building energy at 52.8%, infrastructure maintenance at 17.4%, autonomous robotics at 11.7%, scheduling at 6.3%, safety monitoring at 3.7% and structural design and materials at 3.2%. Output expanded from 2020 onward, representing 92.7% of the corpus, led by policy-gradient and actor-critic architectures. Across all fronts, 85.7% of studies remain in simulation, 12.3% achieve controlled testing and 2.0% achieve field demonstration, with 20 of 24 trials in building operations.
Facility managers can evaluate RL controllers against operational metering data where field validation exists. For earthmoving machinery, crane operations, dynamic scheduling and hazard prevention, contractors can implement digital-twin testbeds and shadow-mode trials to evaluate control policies before physical jobsite deployment.
This research delivers a lifecycle-wide empirical framework that evaluates algorithmic profiles directly against physical deployment environments, separating mature operational applications from simulation-bound methods across the built environment.
