WaveVerse-Sim: Scalable RF Simulation in Generative 4D Worlds

WaveVerse-Sim is the GPU-accelerated RF simulator accompanying Scalable RF Simulation in Generative 4D Worlds. It models spatial and temporal phase coherence and generates FMCW radar signals from indoor scenes and moving human meshes using GPU ray tracing, CUDA signal synthesis, and grouped human scattering.

🛠️ Installation

Simulation requires an NVIDIA GPU. The examples below have been validated on an RTX 3090 Ti with 24 GB VRAM and CUDA 12.2.

git clone git@github.com:penn-waves-lab/WaveVerse-Sim.git
cd WaveVerse-Sim
conda env create -f environment.yml
conda activate waveverse-sim

📦 Example data

Both examples use the bundled laundry room. To make temporal phase coherence easy to understand, the temporal example adds a posed human mesh and moves it back and forth at runtime. This is a simple translation of the entire mesh, not a recorded human-motion sequence or an articulated walking animation.

data/
├── laundry_room/
│   ├── laundry_room.xml
│   ├── meshes/
│   └── scene.json
└── person/
    ├── pose.ply
    └── correspondence.pkl

🚀 Spatial phase coherence

Simulate panoramic high-resolution imaging as described in our paper, and compare the results with and without spatial phase coherence:

python examples/spatial_coherence.py --config configs/spatial.json

With the bundled config, signals and range images are saved to results/spatial/. Open results/spatial/spatial_comparison.png to compare imaging with and without spatial phase coherence.

🚀 Temporal phase coherence

Compare Doppler with and without temporal phase coherence over 0.1–0.8 m:

python examples/temporal_coherence.py --config configs/temporal.json

With the bundled config, per-frame signals are saved to results/temporal/. Open results/temporal/doppler_comparison.png to compare Doppler with and without temporal phase coherence over 0.1–0.8 m.

To plot full-range 0–9.6 m Doppler from the saved signals:

python examples/plot_doppler.py results/temporal --full-range

The full-range figure is saved to results/temporal/doppler_full_range.png.

🙏 Acknowledgements

The codebase is built on Sionna, Mitsuba, Dr.Jit, TensorFlow and CuPy.

📜 Citation

@inproceedings{zheng2026scalable,
  title={Scalable RF Simulation in Generative 4D Worlds},
  author={Zhiwei Zheng and Dongyin Hu and Mingmin Zhao},
  booktitle={Forty-third International Conference on Machine Learning},
  year={2026},
}

License

Code is released under Apache-2.0. Bundled third-party software retains its notices.