An efficient and uncertainty-aware reinforcement learning framework for quality assurance in extrusion additive manufacturing

Defects in extrusion additive manufacturing remain common despite its prevalent use. While numerous AI-based quality assurance approaches have been proposed, the dynamic nature of printing processes often causes deterministic models to lose robustness and, in some cases, fail entirely in new or slightly altered environments. This work introduces an agent that adjusts flow rate and temperature in real-time to optimize control while addressing bottlenecks in training efficiency and uncertainty management. A vision-based uncertainty quantification module generates probabilistic distributions from classified extrusion states, which are integrated with a deep Q-learning controller. While the underlying networks are deterministic, the evolving distributions introduce adaptability to the decision-making process. The controller learns optimal asynchronous actions in a simulation calibrated to vision accuracy and trained with progressively tightened elliptically shaped rewards that account for parameter coupling. With zero-shot learning, the agent bridges the sim-to-real gap and reliably corrects 21 tests across three extrusion error levels—slight, moderate, and severe—with average convergence steps of 40 . 67 ± 17 . 41 , 44 . 00 ± 13 . 56 , and 49 . 11 ± 17 . 91 , respectively. The modest increase in convergence steps and stable standard deviations across error levels underscore the controller’s effectiveness and robustness. Beyond extrusion, this scalable framework supports practical AI-driven quality assurance across various additive manufacturing. • Data-efficient, structured training converges with minor sim-to-real gap. • Inject quantified printing uncertainty into the decision-making process. • Dynamic vision-based distributions enhance the policy network explainability. • Scalable reward design accounts for process parameter coupling effects.