Group Equivariant Convolutional Networks for Pathloss Estimation

TL;DR

The paper tackles pathloss estimation for 6G by introducing RadioGUNet, a UNet enhanced with group equivariant convolutions to capture rotations and reflections without extra data augmentation. The approach enforces symmetry-aware inductive biases by using discrete groups (e.g., C2, C4, C8, D2, D4, D8) and maintains the same input/output setup as RadioUNet, but with a single UNet instead of a dual cascade. Empirical results on the RadioMapSeer dataset show consistent RMSE/NMSE improvements over the baseline across multiple simulation settings, with higher-order groups delivering further gains while maintaining similar parameter counts. The work suggests that symmetry-aware architectures can enhance wireless channel modeling and opens avenues for continuous-symmetry extensions and applications to other radio tasks. Overall, RadioGUNet demonstrates that group equivariant learning can improve pixel-level pathloss estimation with practical computational trade-offs and robust generalization potential.

Abstract

This paper presents RadioGUNet, a UNet-based deep learning framework for pathloss estimation in wireless communication. Unlike other frameworks, it leverages group equivariant convolutional networks, which are known to increase the expressive capacity of a neural network by allowing the model to generalize to further classes of symmetries, such as rotations and reflections, without the need for data augmentation or data pre-processing. The results of this work are twofold. First, we show that typical UNet-based convolutional models can be easily extended to support group equivariant convolution (g-conv). Secondly, we show that the task of pathloss estimation benefits from such an extension, as the proposed extended model outperforms typical UNet-based models by up to 0.41 dB for a similar number of parameters in the RadioMapSeer dataset. The code is publicly available on the GitHub page: https://github.com/EricssonResearch/radiogunet

Figures (2)

• Figure 1: Original UNet architecture (adapted from ronneberger2015u (left) and enhanced UNet with G-Conv (right)
• Figure 2: Example visualized pathloss maps from the three simulation settings in RadioMapSeer. Buildings are in black, and lighter pixels indicate lower pathloss.