Multi-modal deep reinforcement learning for traffic-aware adaptive routing in IoT logistics systems

The growth of IoT-enabled logistics has created rich real-time sensor streams that can support smarter delivery prediction, SLA monitoring, and adaptive routing. This paper presents MARLIN, an end-to-end traffic-aware logistics framework that learns directly from raw multi-modal data, including traffic sequences, LiDAR, camera inputs, IoT telemetry, and weather signals. MARLIN combines reconstruction-based modality encoders, hierarchical cross-attention fusion with learnable gating, an SLA-aware prediction head, and a graph-based Soft Actor-Critic routing agent. Using a controlled benchmark built from NYC Taxi, NGSIM, Open-Meteo, and Chicago Traffic data, MARLIN achieves MAE = 0.797 min, RMSE = 1.101 min, R2 = 0.586, MAPE = 7.88%, ±2-min accuracy = 91.8%, and 99.8% SLA compliance. These results were obtained in a controlled benchmark in which public datasets supplied real traffic, trip, and weather inputs, while delivery-delay targets were generated using a physics-informed model. The findings therefore demonstrate performance under a known generative process and should not be interpreted as validation on observed dispatch-to-delivery outcomes. External validity must be established using co-registered operational logistics data with directly observed delivery delays.