RadioAR: Autoregressive Modeling for Accurate Radio Map Estimation

Peilin Zheng, Xiucheng Wang, Wenshuo Chen, Ruijin Sun, Lei Wang, Yutao Yue, Nan Cheng

Keywords: radio map estimation, autoregressive modeling, multi-scale tokenization, transformer, wireless propagation

TL;DR: RadioAR estimates radio maps by tokenizing them into continuous multi-scale Laplacian-pyramid tokens and autoregressively refining tokens from coarse to fine. It avoids quantization and iterative sampling for fast, accurate reconstruction.

Abstract:

Radio map estimation (RME) provides spatially continuous predic-
tions of radio-frequency metrics such as received signal strength
(RSS) and channel gain, enabling environment-aware operation
in emerging 6G networks. Accurate RME is challenging in com-
plex scenes where multi-path and shadowing create radio maps
that contain both smooth large-scale trends and rapid small-scale
fluctuations. We present RadioAR, a multi-scale autoregressive
framework that generates radio maps progressively from coarse
to fine resolutions. RadioAR introduces a continuous-token multi-
scale tokenizer based on residual Laplacian pyramid decomposition,
which avoids discretization artifacts and preserves fine-grained
variations in continuous-valued radio maps. Experiments on the
RadioMapSeer benchmark show that RadioAR improves estimation
accuracy over representative CNN-, transformer-, and diffusion-
based baselines while reducing inference cost, supporting low-
latency radio cartography.

Submission Number: 12