AssemPlanner: a multi-agent based task planning framework for flexible assembly system

Annals of Forest Science · Springer Nature · 2026-11-13 · DOI: 10.1007/s10845-026-03017-9

Аннотация

In flexible assembly systems, existing task planning methods require a time-consuming configuration process by multiple experts to establish a production line for a new product. To address this challenge, we propose a multi-agent based task planning framework for flexible assembly systems, denoted as AssemPlanner. It takes natural language described tasks as input, which are converted into actionable sequential production operations. It consists of three specialized agents, i.e. SchedAgent, KnowledgeAgent, LineBalanceAgent, and a scene graph. SchedAgent serves as the central reasoning engine, following a Reason-Action-Observe paradigm. It refines plans using process constraints and feedback from the other two agents and the scene graph. LineBalanceAgent places greater emphasis on rapid feasibility and deployment efficiency. In experiments, KnowledgeAgent achieves high accuracy on single-hop questions and multi-hop reasoning, supporting knowledge-grounded reasoning in the evaluated industrial QA tasks. LineBalanceAgent achieves competitive scheduling outcomes while eliminating the need for expert-level parameter tuning and complex reward engineering, facilitating streamlined task adaptation through textual prompts. SchedAgent achieves over 68% accuracy in task planning and over 96% in subtask decomposition. A physical pressure-reducing valve case demonstrates execution feasibility; across four evaluated valve products, program generation took 32–86 s versus 10–25 min for manual programming. To facilitate reproducibility, all code and datasets are released at https://github.com/chz332/Assemplanner .

Читать у издателя →

Похожие статьи