Abstract:

To improve trajectory prediction accuracy, collision risk identification, and probabilistic assessment reliability in industrial warehouse scenarios, a RAPTP method is proposed. RAPTP employs an SE(2)-equivariant state-space encoder to model multi-agent interactions, combines flow matching with a neural hazard field to jointly generate future trajectories and estimate collision risk, and integrates distributionally robust optimization (DRO)with Mondrian conformal prediction for probability calibration. An IWS benchmark is also constructed, followed by comparative, ablation, and zero-shot transfer experiments. RAPTP achieves a minimum Final Displacement Error(minFDE) of 1.50 m on nuScenes, 16.2% lower than EqMotion. On the IWS benchmark, the near-miss detection Area Under the receiver operating characteristic Curve(AUC) is 34.4% higher than Time-to-Collision(TTC) baseline, the early warning time increases from 1.5 s to 3.8 s, and the prediction interval coverage reaches 94.2% at a target coverage level of 90%. In occlusion scenarios of the Hubei dataset, the detection AUC reaches 89.2%. These results demonstrate that RAPTP achieves effective trajectory prediction, collision warning, probability calibration, and cross-scenario transferability.

Key words: industrial warehouse safety (IWS), risk-aware probabilistic trajectory prediction(RAPTP), flow matching, neural hazard field, collision risk