Cross-Modal Graphormer-CRF Prediction Model for Ramp Passage of Warehouse Robots

Source abstract

Warehouse robots frequently traverse short indoor ramps that connect receiving docks, sorting areas and storage floors. Ramp passage is not a routine navigation event because safety depends on slope, payload distribution, wheel slip, acceleration, battery condition, floor friction and route command. This paper proposes a Cross-Modal Graphormer-CRF prediction model for warehouse robot ramp passage. The model aligns inertial measurements, wheel encoder signals, motor current, depth images, load-cell readings and route-state data into passage windows. Cross-modal graph nodes are then constructed to represent ramp geometry, robot attitude, drive response, payload condition and sensing quality. A Graphormer encoder captures physical relations among these evidence groups, while a CRF decoding layer preserves the transition logic of approach, climb, boundary risk and passage failure. A simulated warehouse dataset of 3,210 ramp- passage windows is used for evaluation. The proposed model achieves 91.8% accuracy, 88.6% macro-F1 and 84.3% recall for boundary-risk windows, outperforming LSTM, Transformer and GCN baselines. Robustness tests with missing depth frames, load shift and encoder noise show that confidence decreases before the class label becomes unstable. These results indicate that structured cross-modal modelling can support reliable ramp-passage prediction and provide interpretable evidence for warehouse robot scheduling, route restriction and manual inspection.