| Overview

To alleviate the heavy computational load and the inflexible decision process characteristic of centralized systems, reconfigurable manufacturing can rely on the collaborative teaming of AI-powered modular machines to monitor the production process, communicate, and make decentralized individual decisions. In this work, a decentralized knowledge-based framework is proposed to create the digital twins of collaborative modular machines in reconfigurable manufacturing systems and perform multi-stage dynamic task planning considering task uncertainties, heterogeneity in modular specialization, and reconfiguration management. By taking advantage of simulations that ensure the reproducibility of complex scenarios and data accessibility, decentralized deep reinforcement learning trains the machines to efficiently collaborate on tasks that have sequential constraints, require co-execution, or can be done in parallel. Results demonstrate that the proposed method with multi-objective reward can achieve efficient module usage without compromising completion speed.