Reinforcement Learning for Adaptive Control of Robotic Construction Processes Under Variable Material Conditions

Abstract

Robotic construction processes operate under material conditions that shift with mixing, curing, moisture, temperature, delivery pressure, and layer geometry. A control policy tuned for a nominal mortar or concrete rheology can lose deposition accuracy or violate process limits when yield stress, stiffness, and flow resistance change during execution. This manuscript develops a methodological framework for adaptive control of robotic construction processes by coupling online material-state estimation with safety-constrained reinforcement learning. The robot is modeled as acting on a belief over latent material and process parameters rather than on a single nominal state. A recursive estimator updates those parameters from force, torque, pressure, vision, and kinematic residuals, while the policy learns continuous tool-motion and deposition commands under reward terms for geometric quality, productivity, actuator effort, and constraint violation. Safety is represented through constrained policy optimization with explicit limits on layer stability, collision margin, motor torque, and surface tolerance. A sim-to-real discrepancy bound is used to structure calibration and domain randomization, and uncertainty-sensitive action selection is formulated to avoid aggressive adaptation when material identification is weak. The framework clarifies how estimator covariance, constraint thresholds, reward weights, and simulation fidelity jointly determine adaptation behavior. Its central contribution is an integrated control logic that treats material variability as a partially observable, safety-relevant state rather than as an exogenous disturbance. The analysis provides design criteria for robotic construction systems that must remain productive without sacrificing process stability.