U6G XL-MIMO Radiomap Prediction: Multi-config Dataset & Beam Map Approach

March 13, 2026 · View on GitHub

A public benchmark and reproducibility package for multi-configuration radiomap prediction in U6G / XL-MIMO systems.

Public release: The dataset, pretrained models, source code, and project website are now publicly available.

Overview

This project provides a unified benchmark for radiomap prediction under multiple transmitter configurations in U6G / XL-MIMO systems. It is designed to support reproducible research on:

  • multi-configuration radiomap prediction
  • cross-configuration generalization
  • cross-environment generalization
  • beam-aware radiomap modeling
  • sparse radiomap reconstruction

A key feature of the release is the joint organization of:

  • height maps
  • configuration-aware beam maps
  • ray-tracing radiomap labels
  • ray-tracing scenes and related assets
  • UNet / GAN baseline pipelines

Dataset Preview

Representative examples from the released U6G XL-MIMO multi-configuration dataset, including height maps, ray-tracing radiomaps, and configuration-only beam maps.

Example Galleries

Released Resources

The public release includes:

1. Dataset

  • height maps
  • radiomaps
  • beam maps
  • configuration files
  • ray-tracing scenes and simulation-related assets

2. Baseline Code

  • dataset generation pipeline
  • dataset preprocessing and loading
  • UNet training and evaluation
  • GAN training and evaluation
  • validation and visualization tools

3. Pretrained Models

Pretrained checkpoints for benchmark tasks are released in the Hugging Face repository.

4. Documentation

Detailed documentation is provided on the project website, including dataset structure, benchmark settings, quickstart instructions, and pretrained model organization.

Quick Facts

  • Scenes: 800
  • Frequencies: 1.8 / 2.6 / 3.5 / 4.9 / 6.7 GHz
  • TX antennas: up to 1024 TR
  • Beam counts: 1 / 8 / 16 / 64
  • Beam pattern: 3GPP TR 38.901

Benchmark Tasks

The benchmark uses a unified task naming scheme:

  • random_dense_feature
  • random_dense_encoding
  • beam_dense_feature
  • beam_dense_encoding
  • scene_dense_feature
  • scene_dense_encoding
  • random_sparse_feature_samples819
  • random_sparse_encoding_samples819

These tasks vary in:

  • split strategy: random / beam / scene
  • supervision density: dense / sparse
  • input mode: feature / encoding

Getting Started

Option A: Inspect the released dataset

Download the dataset packages and pretrained models from Hugging Face.

Option B: Evaluate released checkpoints

Use the released pretrained models together with the evaluation scripts in this repository.

Option C: Retrain benchmark models

Use the training scripts together with the documented benchmark task settings.

Option D: Reproduce data generation

Use the dataset-generation scripts to inspect or reproduce the pipeline for scene processing, radiomap generation, validation, and beam-map construction.

For detailed usage instructions and benchmark documentation, please refer to the project website:

Core Dependencies

The released repository spans multiple workflows, and different parts of the codebase rely on different environments.

For evaluation and baseline training

The released UNet / GAN training and evaluation scripts primarily rely on standard Python ML packages such as:

  • PyTorch
  • NumPy
  • Pandas
  • Matplotlib
  • scikit-image

For ray-tracing radiomap generation

The radiomap generation pipeline (DatasetGeneration_Step4_RadiomapRT.py) relies on:

  • TensorFlow
  • Sionna 0.19.2

For scene construction from OSM

The scene-conversion step (DatasetGeneration_Step2_OSMToSionna.py) relies on:

  • Blender 4.0
  • bpy (Blender Python environment)

Important note

Users who only want to evaluate the released pretrained checkpoints do not need to install the full dataset-generation environment.
The Sionna 0.19.2 and Blender 4.0 + bpy dependencies are mainly required for reproducing the released data-generation pipeline.

Citation

If you use this project in your research, please cite:

@misc{li2026u6gxlmimoradiomapprediction,
      title={U6G XL-MIMO Radiomap Prediction: Multi-Config Dataset and Beam Map Approach}, 
      author={Xiaojie Li and Yu Han and Zhizheng Lu and Shi Jin and Chao-Kai Wen},
      year={2026},
      eprint={2603.06401},
      archivePrefix={arXiv},
      primaryClass={eess.SP},
      url={https://arxiv.org/abs/2603.06401}
}

Contact

Xiaojie Li (李宵杰）:xiaojieli@seu.edu.cn/xiaojieli@nuaa.edu.cn