Edge-Adaptive VMamba-UNet for Congestion Prediction in Automated Warehouse Digital Twins

Source abstract

Automated warehouse digital twins require short-horizon forecasts that identify not only whether congestion will occur, but also where a queue front will cross an aisle, buffer, or sorter interface before the next dispatch decision. Existing methods commonly represent synchronized warehouse states as tabular sequences or smooth occupancy maps, which weakens operational boundaries and makes delayed or conflicting sensor observations difficult to handle. This study proposes an edge-adaptive VMamba-UNet for multi-horizon congestion prediction. Camera occupancy, automated guided vehicle telemetry, queue counters, order-release events, route conflicts, and sensor reliability are aligned on a fixed warehouse grid. A pre-encoder fusion module then weights motion discontinuities, topology-defined conflict edges, and confidence changes, while VMamba selective scanning propagates workload pressure across distant zones with linear complexity. The model produces continuous congestion intensity and warning-probability maps at 5, 10, and 15 min horizons. Experiments use a 42-zone simulated and replay-validated warehouse twin containing 2.8 million synchronized records. The proposed model achieves an MAE of 0.061, a warning F1-score of 89.7%, and a boundary F1-score of 83.5%. Compared with Transformer-UNet, MAE decreases from 0.083 to 0.061, warning F1 increases by 7.3 percentage points, and P95 inference latency decreases from 41.8 to 24.6 ms. Removing edge-adaptive fusion increases MAE to 0.070 and reduces boundary F1 to 77.2%. These results indicate that reliability-conditioned boundary preservation and selective state propagation can provide accurate, spatially actionable congestion forecasts within an edge-deployment time budget, although broader multi-site validation remains necessary.