SWRL framework slashes tardiness in assembly scheduling by 30%

A team of scientists from multiple institutions, led by Junhao Qiu, introduces SWRL — a sliding-window reinforcement learning framework designed to tackle the notoriously complex problem of dynamic assembly flow shop scheduling with multi-product kitting delivery. In hybrid manufacturing systems where both processing and assembly happen in sequence, real-time order arrivals create shifting supply dependencies and job-machine assignments. The paper, submitted to arXiv (2607.02941), formalizes this as a heterogeneous graph-based Markov decision process that captures dual-layer kitting structures and tail-product bottlenecks causing sparse rewards.

SWRL's innovation lies in three components: a sliding-window filter that prunes inactive nodes and prioritizes kitting-critical operations, a spatiotemporal graph encoder to track bottleneck migration across consecutive decisions, and a dynamic action mapping module with a constrained waiting strategy that adapts to changing action spaces under variable topologies. Experiments on real-world instances from a home appliance manufacturer show SWRL achieving consistent tardiness reductions over classical dispatching rules and existing deep RL methods. The model remains robust across different resource configurations, order loads, and arrival concentrations, making it a practical solution for smart factories and just-in-time production systems.