Abstract: The final assembly stage of complex products is characterized by deep material hierarchies, strong kitting dependencies, and frequent shop-floor disturbances. Coupled effects of workstation demand changes, automated guided vehicle (AGV) failures, traffic congestion, and line-side inventory constraints make conventional static delivery schedules difficult to execute in a timely manner. To address the material-delivery scheduling problem from the central warehouse and line-side buffers to assembly workstations, this paper proposes a digital twin-based dynamic scheduling method using a constraint-guided deep Q-network (DT-DQN). First, a multi-period mixed-integer linear programming (MILP) model is formulated considering warehouse inventory, line-side inventory, critical material kitting, vehicle capacity, vehicle availability, and dynamic travel time. Delivery task variables are separated from material quantity variables to avoid repeated calculation of travel time when multiple materials are delivered by the same vehicle. Second, the digital twin continuously calibrates material demand, inventory levels, AGV availability, and travel time using real-time execution data. The updated system state is then used to determine the model parameters and feasible action set for the next decision period. Subsequently, atomic action sequences are generated within each decision period through workstation urgency ranking, candidate supply node screening, action masking, and reward shaping. Simulation experiments based on a complex equipment final assembly workshop show that, compared with Rolling-MILP, DT-DQN reduces the weighted production waiting time by 7.1% and improves the solution feasibility rate by 4.9 percentage points under the comprehensive disturbance scenario. Compared with Dispatching Rules, DT-DQN improves the material kitting rate by 9.7 percentage points. Under scaled-up problem instances, DT-DQN exhibits strong online responsiveness, achieving a favorable balance between solution quality and constraint-preservation capability. The proposed method provides theoretical support and practical reference for dynamic scheduling and closed-loop decision-making in complex product final assembly logistics.