MM-LSCM — Multi-Modal Neural Radio Radiance Field

Authors

Who built it

Yiheng Wang Shutao Zhang Ye Xue Tsung-Hui Chang

Shenzhen Research Institute of Big Data / The Chinese University of Hong Kong, Shenzhen · IEEE Globecom Workshops 2025

Why

RSRP alone is blind to geometry

Classical LSCM lacks environmental priors. Pure neural radio fields often need labeled APS and ignore scene structure. MM-LSCM closes both gaps with a dual-branch neural radio radiance field.

Self-supervised radio loss

Train with ŷ = Φ x̂ from beam RSRP—no ground-truth APS required.

LiDAR as a second modality

Voxel density guides hitting probability and depth supervision.

Shared stop probability

One coherence weight renders channel APS and obstacle depth together.

Method

Volume-rendering multi-modal synthesis

Rays leave the BS; each sample queries F_Θ for density σ and direction-dependent signal S. Stop weights synthesize APS x̂ and depth ẑ.

MM-LSCM volume rendering framework — Framework: voxel position, point-cloud density, target grid, and ray direction map to radio and geometry properties.

Results

Lower error, stronger robustness

Indoor laboratory setting. MAE in dB; lower is better.

Method	Sub-Task 1	Sub-Task 2	Sub-Task 3
MM-LSCM	0.87	0.56	0.90
SM-LSCM	1.09	0.70	1.10
WNOMP	1.85	—	—

Sub-Task 1: rotated RSRP @ explored · Sub-Task 2: RSRP @ unexplored · Sub-Task 3: rotated RSRP @ unexplored.

Cite

BibTeX

@inproceedings{wang2025multi,
  title={Multi-modal neural radio radiance field for localized statistical channel modelling},
  author={Wang, Yiheng and Zhang, Shutao and Xue, Ye and Chang, Tsung-Hui},
  booktitle={2025 IEEE Globecom Workshops (GC Wkshps)},
  pages={1005--1010},
  year={2025},
  organization={IEEE}
}