LLM-Guided Hybrid Scheduling With Digital Twin Feedback for Heterogeneous Multi-Robot Aircraft Stringer Assembly

Abstract

The lack of deterministic guarantees under stringent physical constraints, combined with the curse of dimensionality, hinders the deployment of heterogeneous multi-robot systems (MRSs) in aircraft stringer assembly. To bridge the gap between high-level reasoning and low-level feasibility, this article proposes a hybrid scheduling architecture that synergizes large language models (LLMs), mixed-integer linear programming (MILP), and digital twins (DTs). Specifically, we leverage the semantic reasoning of LLMs to extract high-quality prior knowledge, which is then mapped by an adaptive selector to guide the MILP solver through three distinct refinement strategies: filter, warm, and hybrid. This approach transforms LLM-generated insights into kinematically feasible schedules. To enhance operational robustness, a DT-driven event-triggered mechanism is integrated to enable agile rescheduling upon detecting disturbances. Experimental results on FJSP benchmarks and an industrial instance (with up to 64 tasks) show that the proposed method improves the solving speed of the traditional MILP solver on large-scale instances and reliably obtains feasible schedules. Furthermore, the proposed method maintains an average dynamic recovery rate of approximately 80% under random perturbations. This article provides a scalable and resilient solution for complex multi-robot collaboration in high-precision aerospace manufacturing.