A knowledge-lite reinforcement learning framework for online scheduling in human–robot collaborative customized production systems
提出轻知识强化学习框架，通过层次化多网络架构直接学习调度动作，联合优化完工时间、延迟、机器人利用率与人机交互次数，在动态定制化生产中优于传统规则与现有强化学习方法。
To accommodate the increasing demand for personalized customization, manufacturing enterprises must enhance the flexibility of their production lines while efficiently allocating limited robotic resources under strict time constraints. This study tackles the multi-objective scheduling challenges inherent in dynamic, human–robot collaborative (HRC), and highly customized production environments by introducing a knowledge-lite reinforcement learning (KLRL) framework. Unlike traditional rule-based approaches, the KLRL framework employs a hierarchical multi-network architecture in which job and robot features are encoded as state matrices, and scheduling actions are directly derived from the mapping relationships among production factors – thus eliminating the dependence on predefined dispatching rules. The method explicitly incorporates dynamic factors, including stochastic job arrivals and cancelations, scheduled maintenance, robot breakdowns, and human reconfiguration of robots to adapt to new manufacturing processes. Multiple objectives, including makespan, job tardiness, robot utilization, and number of human–robot interactions, are jointly optimized through a dynamic learning mechanism. Extensive experiments across various production scales demonstrate that KLRL delivers superior or comparable performance to composite dispatching rules and state-of-the-art reinforcement learning (RL) methods, thereby validating its effectiveness in complex, human–robot collaborative, and customized scheduling scenarios.