Integration of reinforcement learning in robotic additive manufacturing control: advances, challenges, and future perspectives

Reinforcement learning (RL) is emerging as a promising route to adaptive, data-driven control in robotic additive manufacturing (AM). This review surveys the literature on both traditional and RL-based controllers in AM and finds that, in the case of RL applications, most studies prioritise toolpath optimisation or process-parameter tuning over robotic feedback control, such as motion-platform actuation or power-source regulation, and that the majority remain confined to simulation studies. Only a small, though growing, subset attempts closed-loop control of melt-pool dynamics, bead geometry, or thermal profiles—objectives that are central to the production of high-quality AM parts. Where evaluated, RL demonstrates the ability to exploit high-dimensional sensing to manage nonlinear, multivariable interactions, thereby improving tracking performance, robustness, and part quality. Following an in-depth examination of the state of the art in both traditional controllers and RL applications, this survey identifies and analyses the main barriers to industrial deployment, including the lack of formal stability and safety guarantees, sim-to-real mismatch and sample inefficiency, closed vendor platforms with limited low-latency actuation, and millisecond-scale real-time constraints on edge hardware. Finally, potential solutions to these challenges are discussed, including hybrid RL-traditional control architectures, data-efficient learning with digital twins and reduced-order models, offline RL and human-informed initialisation to minimise on-machine exploration, explainable RL for policy transparency, and hierarchical integration that incorporates additional Artificial Intelligence (AI)-based software modules such as in situ monitoring and anomaly detection to move beyond set-point regulation. Exploring these solutions could harness the most advanced AI techniques in manufacturing to improve learning efficiency, enhance the quality and reliability of feedback controllers, and increase policy interpretability for certification purposes in AM.