Digital twin-based self-learning decision-making framework for industrial robots in manufacturing

Abstract

Industry 4.0 demands intelligent and autonomous manufacturing systems that can adapt to dynamic environments. A key enabler of such systems is self-learning, which supports effective decision-making under uncertainty. However, traditional approaches typically depend on large volumes of physical production data, limiting their utility during early deployment phases. To address this, we propose a digital twin-based self-learning decision-making framework that enables virtual training of decision models before physical deployment. The framework consists of four modular components: a digital twin for simulating the production environment, a learning module for generating and evaluating decisions, a real controller for executing validated actions, and the physical production system. A time-delayed deployment mechanism is introduced to ensure the safe application of learned behaviors. We also define the concept of self-learning in manufacturing and classify relevant learning paradigms—model-driven, data-driven, and hybrid. The framework is validated through two industrial use cases involving FANUC robots. In the first use case, Bayesian optimization is applied to a pick-and-place task, reducing energy consumption by 74.79%. In the second, a Genetic Algorithm is used for welding optimization, achieving a 36.06% reduction in energy usage. These results confirm the framework’s generalizability and effectiveness in enabling autonomous learning during the commissioning phase.