Controllable Radar Simulation with Waveform Parameter Embedding

This repo contains the official code of our paper: Controllable Radar Simulation with Waveform Parameter Embedding.

Authors: Weiqing Xiao^*, Hao Huang^*, Chonghao Zhong^*, Yujie Lin, Nan Wang, Xiaoxue Chen, Zhaoxi Chen, Saining Zhang, Shuocheng Yang, Pierre Merriaux, Lei Lei, Hao Zhao^†

✨ News

• June 4, 2025: Release paper.

• June 3, 2025: Release checkpoints of ICFAR-Net and downstream models.

• June 1, 2025: Release project page.

📝 TODO List

• Release Camera-based pipeline.

📊 Overview

(a) SA-Radar (i.e., Ctrl-RS) enables controllable and realistic radar simulation by conditioning on customizable radar attributes. It supports flexible scene editing such as attribute modification, actor removal, and novel trajectories. (b) SA-Radar improves performance on various tasks including semantic segmentation, 2D/3D object detection. In all settings, SA-Radar’s synthetic data either matches or surpasses real data, and provides consistent gains when combined with real-world datasets.

---

• From left to right are the RGB image and the range-azimuth slices of the radar cube for Ground-Truth, Simulation, Attribute Modification, Actor Removal, and Novel Trajectory.

• Comparison of detection results between the model trained on SA-Radar simulation data and the baseline on real sequences.

• Comparison of detection results between the model trained on SA-Radar simulation data and the baseline on simulated sequences.

🚀 Table of Contents

• Environment Setup
• Dataset Preparation
• Evaluate the Pre-train Model
• Train Your Model
• Run Radar Simulation
• Downstream Tasks

Environment Setup

First, create a new Conda environment and specify the Python version:

conda create -n SA_Radar python=3.11.9
conda activate SA_Radar
pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu121
pip install opencv-python
pip install scikit-image
pip install tensorboard==2.12.0
pip install matplotlib
pip install tqdm
pip install timm==0.5.4
pip install numpy==1.26.4

Dataset Preparation

Please keep the same directory tree as shown in RADDet dataset.

Download the dataset and arrange it as the folloing directory tree,

|-- train
	|-- RAD
		|-- part1
			|-- ******.npy
			|-- ******.npy
		|-- part2
			|-- ******.npy
			|-- ******.npy
	|-- gt
		|-- part1
			|-- ******.pickle
			|-- ******.pickle
		|-- part2
			|-- ******.pickle
			|-- ******.pickle
	|-- stereo_image
		|-- part1
			|-- ******.jpg
			|-- ******.jpg
		|-- part2
			|-- ******.jpg
			|-- ******.jpg
|-- test
	|-- RAD
		|-- part1
            |-- ***
	|-- gt
		|-- part1
            |-- ***
	|-- ***

Evaluate the pre-train models

python evaluate.py --restore_ckpt ./models/icfar-net.pth --attribute
or
python evaluate.py --restore_ckpt ./models/icfar-net_trained_by_A.pth --attribute
or
python evaluate.py --restore_ckpt ./models/icfar-net_trained_by_B.pth --attribute
or
python evaluate.py --restore_ckpt ./models/icfar-net_trained_by_C.pth --attribute

Train your model

prepare for the mixed dataset

python make_mixed_dataset_step1.py 
python make_mixed_dataset_step2.py 

train and eval your model on the mixed dataset

python train.py --logdir ./checkpoints/icfar_mixed_bs3_lr0.0002_50e --train_datasets raddet carrada raddet_by_mr --attribute --segment_mask_loss --l1_loss --sml1_loss
python evaluate.py --restore_ckpt ./checkpoints/icfar_mixed_bs3_lr0.0002_50e/icfar-net.pth --attribute

Run-radar-simulation

run radar simulation on RADDet train set

python demo.py --restore_ckpt ./models/icfar-net.pth --save_numpy --version train --attribute
or 
python demo.py --restore_ckpt ./checkpoints/icfar_mixed_bs3_lr0.0002_50e/icfar-net.pth --save_numpy --version train --attribute

run radar simulation on NuScene v1 mini

python demo_on_NuScenes.py --restore_ckpt ./models/icfar-net.pth --attribute --time_steps 100
or
python demo_on_NuScenes.py --restore_ckpt ./models/icfar-net_wo-re.pth --time_steps 100

Scene Editing

Attribute Modification

Modify the attribute_list in demo.py directly.

Novel Trajectories

python demo.py --restore_ckpt ./models/icfar-net.pth --save_numpy --version train --attribute --angle_rotation

Actor Removal

python demo.py --restore_ckpt ./models/icfar-net.pth --save_numpy --version train --attribute --remove

Downstream Tasks

We provide pre-trained weights of the models on different downstream tasks in the downstream_ckps folder, including real-data-trained, simulated-data-trained, and co-trained versions.

Running a Pre-trained Model

For 3D detection, copy the checkpoints from ./downstream_ckps/3d-det folder into RADDet_Pytorch and run:

# RAD head
python validate.py --config_dir ./configs/config.json --resume_from {model.pth}
or
# RA head
python validate_cart.py --config_dir ./configs/config.json --resume_from {model.pth}

For 2D detection (RD), copy the config and checkpoint files from ./downstream_ckps/2d-det (RTMDet Model on mmdet) folder into mmdetection, and run:

python tools/test.py --config configs/{config.py} --checkpoint work_dirs/{model.pth}

Training Your Downstream Model

After running demo.py, the generated simulation data is saved to the ./sim_output folder (set via --output_directory). The simulation data format is identical to the RADDet dataset, which you can use to train your downstream model.

🤝 Citation

If you find this repository helpful, please consider citing our paper:

@InProceedings{Xiao_2026_CVPR,
    author    = {Xiao, Weiqing and Huang, Hao and Zhong, Chonghao and Lin, Yujie and Wang, Nan and Chen, Xiaoxue and Chen, Zhaoxi and Zhang, Saining and Yang, Shuocheng and Merriaux, Pierre and Lei, Lei and Zhao, Hao},
    title     = {Controllable Radar Simulation with Waveform Parameter Embedding},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings},
    month     = {June},
    year      = {2026},
    pages     = {6424-6434}
}