LLM-Assisted Warm-Start Optimization for Multi-Factory Human-Robot Collaborative Disassembly Scheduling

Abstract

The quality of initial solutions strongly affects the convergence behavior of meta-heuristic algorithms for human-robot collaborative disassembly scheduling. This paper proposes a Large Language Model (LLM)-assisted warm-start method for generating feasible task sequences in a multi-factory human-robot collaborative disassembly problem. A structured instruction set is designed to describe product disassembly information and task constraints, including precedence and conflict relationships derived from AND/OR graphs. The generated sequences are parsed, validated, and injected into the initial population of an Intelligent Discrete Optimization Algorithm (IDOA). A profit-oriented mathematical model is used as the evaluation basis for task-tool assignment, workstation allocation, and disassembler configuration. Numerical experiments on several end-of-life (EoL) products show that the proposed LLM-assisted initialization method improves initial solution quality and accelerates convergence in terms of iteration count. Although local LLM inference introduces additional runtime overhead, the results demonstrate the feasibility of using LLMs as warm-start generators for combinatorial disassembly optimization.