Multitask fine-tuning agentic AI based collaborative scheduling for flexible manufacturing systems
Document Type
Research-Article
Journal Name
Journal of Manufacturing Systems
Keywords
Collaborative scheduling, Flexible manufacturing system, Large language model, Multi-agents system, Multitask fine-tuning
Abstract
In the dawn of Industry 5.0, flexible manufacturing systems (FMSs) need to operate under multi-source information, heterogeneous resources, and dynamic environments, thus traditional approaches can hardly achieve adaptability, robustness and interpretability in cross-task collaborative scheduling. To address this challenge in collaborative scheduling of FMS (CSFMS), this paper proposes a framework that comprises large language model (LLM)-based orchestrator-workers agents (OWAs) to perform multi-objectives CSFMS, including heterogeneous manufacturing resource management, multitask orchestration, and multi-process collaboration. To efficiently manage heterogeneous manufacturing resources, an OWA consisting of a planning-level agent and execution-level agents, which are deployed on the dispatching systems of machines and automated guided vehicles (AGVs), and augmented reality (AR) equipped by workers, is developed. To enhance cross-task (i.e. FMS knowledge mapping, agent skills for algorithms, and semantic reasoning and decision making) orchestration capabilities and reduce parameters interference across multitask of LLM, a data generation method based on meta-framework is introduced, as well as a multitask fine-tuning (MFT) with core parameters freezing and drop and rescale (DARE). Moreover, an LLM-based OWA (LLM-OWA) is designed with perception and self-coordination to close-loop collaborative decision. The enhancement of MFT strategies for integrated scheduling LLM (ISLLM) under multi-task scenarios is confirmed by the ablation experiment. For the scheduling results across 33 extended benchmark instances show that the solution space of LLM-OWAs improves stability (with 18.06% standard deviation reduce at least) and quality (with 3.03% average value improvement at least) over baselines (e.g. NSGA-Ⅱ, NSGA-Ⅲ). Under the impact of abnormal events like state failure, task time failure, and emergency order insertion, the proposed method further achieves 0.9146 decision consistency value and 0.9496 interpretability score. This work demonstrates the potential of agentic AI to support next-generation CSFMS. © 2026 The Society of Manufacturing Engineers