AI-Enabled Autonomous Manufacturing
Sungjong Kim1, Chan Hee Park² and Byeng D. Youn1,3,*
1 Department of Mechanical Engineering, Seoul National University, Seoul 08826, Republic of Korea 2 Department of Mechanical and Information Engineering, University of Seoul, Seoul 02556, Republic of Korea 3 Onepredict Corp., Seoul 06105, Republic of Korea * Corresponding author
E-mail: bdyoun@snu.ac.kr
Status
Automated manufacturing refers to the use of control systems, machinery, and information technologies to execute predefined production tasks with minimal human intervention. While such systems have contributed significantly to productivity gains, they still rely heavily on rule-based logic or expert knowledge, which limits their adaptability to dynamic environments and complex manufacturing tasks. Recently, the combination of a declining skilled experts, rising wages and energy costs, and growing demand for high-mix, low-volume production has highlighted the need for transformative innovation in manufacturing systems. Autonomous manufacturing represents an evolutionary step forward. It refers to cyber-physical production systems wherein machines, software agents, and embedded systems independently perform sensing, reasoning, and action using distributed intelligence eliminating the need for human oversight in both routine and unstructured scenarios [1]. By digitizing domain expertise and leveraging large-scale process data, autonomous systems provide scalable and adaptive alternatives. Recent advances in artificial intelligence (AI) have enabled these systems to autonomously incorporate real-time feedback, allowing for predictive quality assurance, anomaly detection, and self-optimization of process parameters. The implementation of AI-driven autonomy has been shown to significantly enhance operational efficiency, reduce overhead costs, and improve system resilience—particularly in globally distributed manufacturing environments where access to expert knowledge is limited. Empirical evidence highlighted the effectiveness of such technologies; for example, the implementation of an autonomous quality management system in the automotive manufacturing sector resulted in a 52% reduction in production costs and a 78% decrease in inspection expenses [2]. Moreover, autonomous manufacturing technologies are expected to exhibit broad applicability across diverse operational domains, including quality control, logistics, energy management, equipment maintenance, and comprehensive process optimization.
Current and future challenges
Achieving truly AI-enabled autonomous manufacturing requires seamless integration of three foundational components—sensing, reasoning, and action—while also establishing a robust platform for managing the integrated autonomous manufacturing system. In the sensing stage, manufacturing systems must establish robust and scalable data pipelines capable of reliably extracting, pre-processing, storing, and managing diverse multimodal sensor data. Despite the abundance of available data, current pipeline architectures are often underdeveloped compared to the overall maturity of production systems. These pipelines are frequently designed without sufficient consideration for downstream reasoning and control tasks. Consequently, the acquired data suffers from data availability issues—such as noise, low resolution, inconsistent sampling, an excessive amount of data and poor synchronization with system context—which hinders the systems’ ability to transmit only relevant, high-quality data necessary for real-time decision-making and autonomous operation [3]. The reasoning stage involves deriving actionable insights support process-level decisions. At this stage, two central challenges arise: ensuring the interpretability and generalization of AI models. For AI systems to contribute effectively to manufacturing operations, they must provide structured information across key categories, including current and predicted system states (system assessment), identified operational tasks (problem definition), causal factors (root cause diagnosis), and prescriptive recommendations (decision-making). However, many AI models operate as “black boxes,” hindering engineers' ability to verify or trust the inferred outputs. Generalization also remains problematic, as models often struggle to maintain robust performance under domain shifts, such as variations in operating conditions, product configurations, or factory environments, leading to physically inconsistent or non-representative results [4]. The action stage requires translating reasoning outputs into executable operations, such as control commands, optimal setpoint selection, or human-readable decision reports. Despite recent progress in AI, current AI models often produce outputs in abstract or model-centric forms that lack the semantic clarity necessary for effective interpretation and implementation within manufacturing systems. Without additional
contextualization, these outputs are not readily actionable, requiring engineers to manually interpret the reasoning results and determine appropriate interventions, thereby increasing cognitive burden and delaying operational response [5]. Lastly, current platforms such as manufacturing execution system (MES), and programmable logic controller (PLC) are hierarchical and lack the flexibility to support autonomous manufacturing operations. Key challenges include poor interoperability across distributed manufacturing components, limited support for real-time self-organization and manufacturing lifecycle integration.
Advances in science and technology to meet challenges
In the sensing stage, data pipelines integrated with extract-transform-load (ETL) mechanisms are employed to convert raw signals into structured, analysis-ready formats [6]. Virtual sensing techniques are utilized to estimate difficult-to-measure variables by leveraging data acquired from the manufacturing process [7]. To enhance data quality and contextual fidelity, pre-processing methods such as noise removal, sampling rate alignment, and synchronization of heterogeneous data sources are applied [8]. Additionally, ontology-based technologies have been developed to define the identity of collected data and establish contextual relationships among correlated information [9]. By enabling context-aware data linkage and semantic interpretation, it facilitates data filtering and selection in subsequent stages, despite the abundance and heterogeneity of manufacturing data. In the reasoning stage, interpretability has been advanced through explainable AI techniques, including pre-modelling strategies such as domain-informed feature extraction, as well as post-modelling tools such as attention mechanism analysis, Shapley additive explanations (SHAP) [10]. To improve generalization under domain shifts, lifecycle-aware learning strategies are employed to support data drift detection and continual learning [11]. Furthermore, efforts to integrate physical constraints into AI architectures—through physics-inspired components (e.g., wavelet kernels) and regularization techniques based on governing equations (e.g., differential constraints)—help ensure physical consistency and reliability across diverse operational settings [12,13]. In the action stage, the primary objective is to translate AI outputs into actionable manufacturing decisions. Natural language interfaces powered by large language models (LLMs) enable the summarization and structuring of outputs into human-readable formats, thereby enhancing interpretability and operational readiness[14]. This requires aligning linguistic representations with manufacturing data to ensure contextual relevance and facilitating the integration of domain expertise through instruction tuning and agent-based LLMs [15,16]. Reinforcement learning (RL)-based optimization methods, including proximal policy optimization (PPO) and deep-Q-networks (DQN), are employed to derive adaptive control strategies from reasoning outputs, allowing systems to respond effectively to dynamic operational conditions[17,18]. Additionally, machine learning operations (MLOps) frameworks support the continuity and reliability of AI-driven actions through version control, performance monitoring, and feedback-based retraining, ensuring sustained robustness across the system lifecycle [19]. Building on these advances, a decentralized autonomous manufacturing (DAM) platform architecture was introduced to enable autonomous decision-making, decentralized control, and self-organizing production capabilities [20]. By utilizing multi-agent systems and secure communication protocols, the platform allows distributed manufacturing nodes to collaborate effectively, respond to disruptions, and execute manufacturing tasks without centralized coordination.
Sensing
Extract Transform Load
ETL process
Raw data Pre-processing Virtual sensing
Autonomous Manufacturing
Platform
Amplitude 2D map Impact on model output Manufacturing AI Phase Features Explainable AI (pre/post modeling) Instruction tuning with LLMs
Physics-based kernel AI model PDE loss Job queue RL Agents DataOps ModelOps
Physics-inspired architecture Physical regularization RL-based scheduling MLOps
Reasoning Action
Figure 1. State-of-the-art research landscape in the three foundational components of AI-enabled
autonomous manufacturing—sensing, reasoning, and action—supported by an integrated autonomous manufacturing platform.
Concluding remarks
AI-enabled autonomous manufacturing is poised to redefine industrial operations by embedding distributed intelligence across the sensing, reasoning, and action layers of production systems. Moving beyond traditional rule-based automation, autonomous systems leverage advanced AI models to make context-aware decisions, adapt to dynamic environments, and self-optimize processes with minimal human intervention. This paradigm shift is increasingly critical in light of global challenges such as declining skilled experts, escalating operational costs, and rising demand for agile, high-mix production. Recent technological advances collectively enable machines and software agents to autonomously perceive, interpret, and act within complex manufacturing settings. Moreover, the emergence of decentralized autonomous manufacturing platforms offers a resilient and scalable infrastructure for self-organizing production systems. By integrating multi-agent systems with secure, scalable communication, these platforms allow distributed manufacturing nodes to collaborate effectively, respond to disruptions in real time, and execute tasks autonomously—without relying on centralized control. The anticipated benefits are far-reaching, encompassing predictive quality control, intelligent maintenance, energy optimization, and logistics coordination. As AI models become increasingly
interpretable, robust, and contextually aware, autonomous manufacturing systems are expected to form the backbone of next-generation smart factories—capable of operating efficiently, responding adaptively, and continuously improving under industrial conditions.
Acknowledgements
This research was partially supported by the International Research & Development Program of the National Research Foundation of Korea (NRF) funded by the Ministry of Science and ICT (No. 2022K1A4A7A04096329), and Technology Innovation Program (or Industrial Strategic Technology Development Program-Automotive Industry Technology Development-Green Car) (RS-2024-00444961, Development and demonstration of Purpose-Built Electric vehicles using design platforms) funded by Ministry of Trade, Industry & Energy (MOTIE, Korea).
References
[1] Mo F, Monetti FM, Torayev A, Rehman HU, Mulet Alberola JA, Rea Minango N, Nguyen HN, Maffei A, Chaplin JC. A maturity model for the autonomy of manufacturing systems. International Journal of Advanced Manufacturing Technology. 2023 May 1;126(1–
2):405–28. [2] Singh J, Ahuja IPS, Singh H, Singh A. Development and Implementation of Autonomous Quality Management System (AQMS) in an Automotive Manufacturing using Quality 4.0 Concept– A Case Study. Comput Ind Eng. 2022 Jun 1;168. [3] Goknil A, Nguyen P, Sen S, Politaki D, Niavis H, Pedersen KJ, Suyuthi A, Anand A, Ziegenbein A. A Systematic Review of Data Quality in CPS and IoT for Industry 4.0. ACM Comput Surv. 2023 Dec 31;55(14 S). [4] Li D, Yang Y, Song YZ, Hospedales TM. Deeper, Broader and Artier Domain Generalization. 2017. [5] Naqvi MR, Elmhadhbi L, Sarkar A, Archimede B, Karray MH. Survey on ontology-based explainable AI in manufacturing. Journal of Intelligent Manufacturing. Springer; 2024. [6] Walha A, Ghozzi F, Gargouri F. Data integration from traditional to big data: main features and comparisons of ETL approaches. Journal of Supercomputing. 2024 Dec 1; [7] Pal A, Kant K. Exploiting Proxy Sensing for Efficient Monitoring of Large-Scale Sensor Networks. ACM Trans Internet Technol. 2020 May 1;20(2). [8] Tawakuli A, Havers B, Gulisano V, Kaiser D, Engel T. Survey:Time-series data preprocessing: A survey and an empirical analysis. Journal of Engineering Research (Kuwait). 2024; [9] Sapel P, Molinas Comet L, Dimitriadis I, Hopmann C, Decker S. A review and classification of manufacturing ontologies. J Intell Manuf. 2024; [10] Kim S, Lee S, Lee J, Kim M, Kim SJ, Yoon H, Youn BD. Fault-relevance-based, multi-sensor information integration framework for fault diagnosis of rotating machineries. Mech Syst Signal Process. 2025 Jun 1;232. [11] Tercan H, Deibert P, Meisen T. Continual learning of neural networks for quality prediction in production using memory aware synapses and weight transfer. J Intell Manuf. 2022 Jan 1;33(1):283–92. [12] Würth T, Krauß C, Zimmerling C, Kärger L. Physics-informed neural networks for data-free surrogate modelling and engineering optimization – An example from composite manufacturing. Mater Des. 2023 Jul 1;231. [13] Li T, Zhao Z, Sun C, Cheng L, Chen X, Yan R, Gao RX. WaveletKernelNet: An Interpretable Deep Neural Network for Industrial Intelligent Diagnosis. IEEE Trans Syst Man Cybern Syst. 2022 Apr 1;52(4):2302–12. [14] Zhang Q, Xu C, Li J, Sun Y, Bao J, Zhang D. LLM-TSFD: An industrial time series human-in-the-loop fault diagnosis method based on a large language model. Expert Syst Appl. 2025 Mar 10;264. [15] Sun Y, Zhang Q, Bao J, Lu Y, Liu S. Empowering digital twins with large language models for global temporal feature learning. J Manuf Syst. 2024 Jun 1;74:83–99. [16] Zheng S, Pan K, Liu J, Chen Y. Empirical study on fine-tuning pre-trained large language models for fault diagnosis of complex systems. Reliab Eng Syst Saf. 2024 Dec 1;252. [17] He Z, Tran KP, Thomassey S, Zeng X, Xu J, Yi C. Multi-objective optimization of the textile manufacturing process using deep-Q-network based multi-agent reinforcement learning. J Manuf Syst. 2022 Jan 1;62:939–49. [18] Zhang Y, Zhu H, Tang D, Zhou T, Gui Y. Dynamic job shop scheduling based on deep reinforcement learning for multi-agent manufacturing systems. Robot Comput Integr Manuf. 2022 Dec 1;78. [19] Bayram F, Ahmed BS. Towards Trustworthy Machine Learning in Production: An Overview of the Robustness in MLOps Approach. ACM Comput Surv. 2025 Jan 22;57(5). [20] Leng J, Zhong Y, Lin Z, Xu K, Mourtzis D, Zhou X, Zheng P, Liu Q, Zhao JL, Shen W. Towards resilience in Industry 5.0: A decentralized autonomous manufacturing paradigm. Vol. 71, Journal of Manufacturing Systems. Elsevier B.V.; 2023. p. 95–114.