Manufacturing AI: From Smart Factories to Sustainable Robotics

Latest 50 papers on manufacturing: Dec. 21, 2025

The world of manufacturing is undergoing a profound transformation, driven by rapid advancements in AI and machine learning. From optimizing production lines and ensuring product quality to enabling more sustainable practices and safer human-robot collaboration, AI is proving to be a game-changer. This blog post dives into recent breakthroughs, distilling key insights from a collection of cutting-edge research papers that showcase the incredible potential and ongoing evolution of AI in manufacturing.

The Big Ideas & Core Innovations

At the heart of these advancements lies a unified effort to make manufacturing processes more intelligent, efficient, and resilient. A central theme is the integration of AI with real-world physical systems, bridging the digital and physical realms. For instance, the “Pretrained Battery Transformer (PBT)” from The Hong Kong University of Science and Technology (Guangzhou) introduces the first foundation model for battery life prediction, leveraging domain-knowledge-encoded mixture-of-expert layers. PBT not only achieves superior accuracy, outperforming existing models by nearly 20%, but also generalizes across diverse battery chemistries, promising a revolution in battery degradation modeling. This highlights a trend towards specialized, physics-informed AI models that understand the nuances of industrial processes.

Similarly, in additive manufacturing (AM), “Closed Loop Reference Optimization for Extrusion Additive Manufacturing” by Guidetti, Fravolini, Moretti, Xia, Wu, Chesser, Bahrami, Kolmanovsky, Zomorodi, and Landers improves extrusion force tracking by up to 39.57% using LQR control and quadratic programming. This significantly reduces material waste and enhances precision, addressing critical challenges in 3D printing. This focus on precision control extends to robotic tasks, where “Iterative Tuning of Nonlinear Model Predictive Control for Robotic Manufacturing Tasks” by Bristow, Tharayil, Alleyne, Rupenyan, Khorsavi, and Lygeros introduces an iterative learning control (ILC)-inspired framework for automatically tuning NMPC weights. This data-driven approach allows robots to rapidly learn and improve performance in repetitive tasks, achieving near-optimal results within just a few repetitions.

Another significant innovation is the emergence of semantic-aware and context-rich AI systems. “Semantic-Constrained Federated Aggregation: Convergence Theory and Privacy-Utility Bounds for Knowledge-Enhanced Distributed Learning” by Jahidul Arafat (Auburn University) introduces SCFA, which embeds domain knowledge into federated learning. This not only accelerates convergence but also improves privacy-utility trade-offs, making distributed learning more robust for sensitive industrial data. Extending this, “Towards Logic-Aware Manipulation: A Knowledge Primitive for VLM-Based Assistants in Smart Manufacturing” by Shridhar, Jiang, Driess, Brohan, Kim, Stone, Nair, Raibert, and Hogan (affiliations including MIT CSAIL and Google Research) proposes a logic-aware schema to enhance vision-language models (VLMs) for precise manipulation. By incorporating structured metadata and explicit affordance structures, VLMs can achieve greater precision and safety in tasks like 3D printer spool removal, bridging the gap between high-level commands and fine-grained robotic actions.

The push for real-time, on-device intelligence is also evident. “Continual Learning at the Edge: An Agnostic IIoT Architecture” by García-Santaclara, Fernández-Castro, Díaz-Redondo, Calvo-Moa, and Mariño-Bodelón (Universidade de Vigo, GRADIANT) develops a modular edge-computing architecture for continual learning in industrial IoT, particularly for real-time quality control. This system efficiently adapts to non-stationary data streams and reduces catastrophic forgetting, proving its value in a cheese production use case. Similarly, “On-Device Continual Learning for Unsupervised Visual Anomaly Detection in Dynamic Manufacturing” by Ren, Köhle, Dorofeev, and Anicic (Technical University of Munich, Siemens AG) enables rapid, memory-efficient adaptation to new product variations directly on edge devices, achieving a 12% AUROC improvement and reducing memory usage by 80%. This is crucial for dynamic manufacturing environments where continuous adaptation is key.

Lastly, the papers highlight an increasing emphasis on sustainability and robust design. “LLM-Upgraded Graph Reinforcement Learning for Carbon-Aware Job Scheduling in Smart Manufacturing” by Yang, Liu, Zhang, Lou, Low, and Gan (Singapore Institute of Technology) introduces Luca, an LLM-upgraded graph reinforcement learning framework that optimizes job scheduling to reduce makespan and carbon emissions simultaneously, demonstrating improvements of up to 12.2% in makespan. This reflects a growing trend toward AI that not only enhances efficiency but also supports environmental goals. Furthermore, “A Physics-Constrained, Design-Driven Methodology for Defect Dataset Generation in Optical Lithography” by Hu, Kong, Shin, Kim, and Kang (KITECH, Hanyang University) uses a physics-driven approach to generate high-fidelity defect datasets for optical lithography, crucial for robust AI-based quality control in semiconductor manufacturing. This method creates pixel-accurate defect datasets, significantly boosting the performance of defect detection models like Mask R-CNN over Faster R-CNN.

Under the Hood: Models, Datasets, & Benchmarks

These innovations are powered by a combination of novel models, carefully curated datasets, and robust benchmarking frameworks:

• PBT (Pretrained Battery Transformer): A domain-knowledge-encoded mixture-of-expert architecture, trained on 13 diverse lithium-ion battery datasets to achieve state-of-the-art life prediction. Code: https://github.com/Ruifeng-Tan/PBT.
• PrediFlow: A flow-based prediction-refinement framework for real-time human motion prediction in human-robot collaboration, demonstrating improved accuracy and faster inference. Resources: https://drive.google.com/file/d/1wV6YYrSEUc1fwO2qRxapiFcWZKkM-LMY/view.
• FiLM-enhanced DeepONet: Developed in “Probabilistic Predictions of Process-Induced Deformation in Carbon/Epoxy Composites Using a Deep Operator Network” by Kiyania et al. (Brown University, University of Delaware, Oak Ridge National Laboratory), this model predicts cure, viscosity, and deformation in composites, integrating physics-based modeling with deep learning for robust PID mitigation. Paper: Probabilistic Predictions of Process-Induced Deformation in Carbon/Epoxy Composites Using a Deep Operator Network.
• PyCAALP: A Python-based framework by Hartmann et al. (Technical University of Munich) that uses graph-based optimization and Mixed-Integer Programming (MIP) for automated assembly sequence and production line planning, incorporating geometric constraints. Code: https://github.com/TUM-utg/PyCAALP.
• DTW-TL Framework: Introduced in “An Additive Manufacturing Part Qualification Framework: Transferring Knowledge of Stress-strain Behaviors from Additively Manufactured Polymers to Metals” by Duan and Wu (University of Central Florida), this framework uses dynamic time warping and transfer learning to predict stress-strain behaviors of AM metals from polymer data, significantly reducing data collection needs. Paper: An Additive Manufacturing Part Qualification Framework: Transferring Knowledge of Stress-strain Behaviors from Additively Manufactured Polymers to Metals.
• VLM-IRIS: A zero-shot framework by Mahjourian and Nguyen (Michigan Technological University) that adapts vision-language models (VLMs) to infrared industrial sensing by converting thermal images into RGB-compatible inputs for CLIP-based encoders. Paper: Vision-Language Models for Infrared Industrial Sensing in Additive Manufacturing Scene Description.
• LLM4SFC: A framework by Glick et al. (Bosch Research) that leverages LLMs to generate executable Sequential Function Charts (SFCs) from natural language, using fine-tuning with retrieval-augmented generation (RAG) and real-time pruning for compliance. Paper: LLM4SFC: Sequential Function Chart Generation via Large Language Models.
• JOCA: An end-to-end framework by Yan, Bryson, and Dansereau (The University of Sydney) that jointly optimizes camera hardware, adaptive control algorithms, and downstream perception tasks, using a hybrid DF-Grad strategy. Code: https://roboticimaging.org/Projects/JOCA/.
• TS-HINT: A novel time series foundation model by Rico, Raghavan, and Jayavelu (SUTD, A*STAR) that integrates LLM reasoning with attention hints to enhance semiconductor manufacturing process prediction. Paper: TS-HINT: Enhancing Semiconductor Time Series Regression Using Attention Hints From Large Language Model Reasoning.
• CORE Metamodel: Introduced by Bertrand et al. (Université de Paris), this object-centric metamodel provides a unified format for IoT-enhanced event logs, addressing interoperability for process mining and data integration. Paper: An object-centric core metamodel for IoT-enhanced event logs.
• SMART+ Framework: Proposed by Kandikatla and Radeljić (MaxisIT Inc., Aula Fellowship for AI), this comprehensive model evaluates and governs AI systems across industries, integrating global AI ethics principles into actionable pillars. Paper: The SMART+ Framework for AI Systems.

Impact & The Road Ahead

The implications of this research are vast, pointing towards an era of highly autonomous, adaptive, and sustainable manufacturing. The ability to predict battery life with greater accuracy, optimize additive manufacturing processes, and enhance human-robot collaboration will lead to reduced waste, improved product quality, and safer working environments. The shift towards edge computing for continual learning signifies a future where AI operates in real-time, directly on the factory floor, adapting to dynamic conditions without constant cloud reliance. This not only speeds up decision-making but also enhances data privacy and reduces latency.

The development of semantic-aware AI, powered by knowledge graphs and large language models, will allow industrial systems to understand and execute complex natural language commands, democratizing access to advanced automation. The focus on robust and ethical AI governance, as highlighted by frameworks like SMART+, ensures that these powerful technologies are developed and deployed responsibly. Furthermore, the push for data-driven sustainable product development, carbon-aware scheduling, and advanced material optimization in areas like heat exchangers and drug delivery systems, demonstrates AI’s critical role in addressing global environmental and health challenges.

Looking ahead, the convergence of these themes promises manufacturing systems that are not only smarter and more efficient but also deeply integrated with human operators, environmentally conscious, and inherently adaptable. The future of manufacturing is intelligent, interconnected, and on the cusp of a new era of innovation.