Deep Learning-Based Cross-Floor Dynamic Channel Modeling and Prediction for Food Delivery Robots

To enhance the reliability of communication systems for indoor service robots in complex cross-floor environments and support the development of smart building IoT and unmanned delivery technologies, this paper conducts a deep learning-based intelligent channel modeling study for multi-floor dynamic mobile scenarios. Addressing the severe Non-Line-of-Sight blockages and strong channel non-stationarity caused by complex indoor topologies and cross-floor movements, this study constructs a full-link dynamic channel dataset at the 5.8 GHz band in a dual-floor building based on high-precision ray-tracing technology, covering five typical physical scenarios: corridors, corners, stairwells, rooms, and floor switching. Through systematic comparison and statistical characteristic analysis, it is found that: compared to traditional empirical models (such as ITU-R P.1238) that assume a monotonically increasing path loss, data-driven models can accurately capture the non-linear fading caused by abrupt topological changes, improving prediction accuracy by over 50%; compared to sequence memory and local perception networks, the global attention architecture (Transformer) exhibits superior generalization robustness in handling complex cross-floor sudden changes; compared to highly complex deep networks, the lightweight Multi-Layer Perceptron model incorporating refined physical feature vectors achieves statistically equivalent prediction performance to Transformer, significantly enhancing parameter utilization efficiency. Accordingly, a deep learning channel modeling mechanism integrating explicit spatial geometric priors is proposed. By forcibly injecting key physical features such as same-floor indicators, floor differences, and cross-floor interactions, the multi-path time-varying evolution and loss jumps caused by complex indoor blockages are accurately characterized. Finally, through multi-seed cross-validation, non-parametric significance testing, and feature ablation comparison, the accuracy and consistency of the statistical characteristics of the established model are verified, providing a reliable theoretical basis and engineering reference for the design and performance evaluation of edge communication systems for computation-constrained mobile robots.