RadarGen: Automotive Radar Point Cloud Generation from Cameras

   

💡 TL;DR: RadarGen generates sparse radar point clouds from multi-view camera images.

🔥 News

• Evaluation code is coming soon.
• [2026/03] RadarGen's inference and training pipelines for MAN TruckScenes and nuScenes are now available.
• [2025/12] Paper is on Arxiv!

🛠️ Installation

Prerequisites

• Python >= 3.10
• CUDA 12.1
• Conda

Clone the Repository

git clone --recursive https://github.com/tomerborreda/RadarGen.git
cd RadarGen

Note: The --recursive flag is required to fetch the UniDepth and UFM submodules. If you already cloned without it, run: git submodule update --init --recursive

Install

bash environment_setup.sh radargen
conda activate radargen

This will install RadarGen and all dependencies including:

• PyTorch 2.4.0 with CUDA 12.1
• UniDepth
• UFM

For manual installation, follow the steps in environment_setup.sh.

Note: A GPU is required for the installation of UniDepth.

Dataset Setup

We support multiple autonomous driving datasets. To get started, download and set up one of the following:

TruckScenes:

• Download: man.eu/truckscenes
• Install devkit: pip install truckscenes-devkit

nuScenes:

• Download: nuscenes.org
• Install devkit: pip install nuscenes-devkit

After downloading, update the configs in configs/*.yaml and configs/preprocessing/*.yaml to point to your dataset location.

🚀 Quick Start

Inference

Run inference on TruckScenes using the provided notebook: notebooks/inference_truckscenes.ipynb

Or use the Python API:

from radargen.inference import RadarGenInference
from radargen.datasets import get_adapter

# Load dataset adapter
adapter = get_adapter("truckscenes", trucksc=trucksc_obj)

# Initialize model
model = RadarGenInference(
    adapter=adapter,
    config_path="configs/RadarGen_600M_512px_TS_inference.yaml",
    checkpoint_path="path/to/checkpoint"
)

# Generate point cloud from two consecutive frames
pcl = model.from_sample_data(sample_t0, sample_t1)

Download pretrained models:

Pretrained model weights are available through huggingface.

• MAN TruckScenes:
  • Link: https://huggingface.co/TomerBo/RadarGen_600M_512px_TS/
  • Path: hf://TomerBo/RadarGen_600M_512px_TS/RadarGen_600M_512px_TS.safetensors
• nuScenes: Coming soon

🏋️ Training

📊 Data Preparation

Before training, you need to create BEV conditioning maps and ground truth radar maps:

1. Create BEV Conditioning Maps

First, edit configs/preprocessing/truckscenes_bev_maps.yaml to set your dataset and output paths.

# Single GPU
python scripts/create_bev_maps.py \
    --config_path configs/preprocessing/truckscenes_bev_maps.yaml

# Multi-GPU (8 GPUs)
bash scripts/run_create_bev_maps.sh 8 \
    --config_path configs/preprocessing/truckscenes_bev_maps.yaml

This creates RGB appearance, semantic segmentation, and velocity maps in bird's eye view.

2. Create Ground Truth Radar Maps

First, edit configs/preprocessing/truckscenes_radar_maps.yaml to set your dataset and output paths.

# Single GPU
python scripts/create_radar_maps.py \
    --config_path configs/preprocessing/truckscenes_radar_maps.yaml

# Multi-GPU (8 GPUs)
bash scripts/run_create_radar_maps.sh 8 \
    --config_path configs/preprocessing/truckscenes_radar_maps.yaml

This creates Point Density, RCS, and Doppler maps from the ground truth radar data.

Training on TruckScenes

1. Prepare the config file

   Edit configs/RadarGen_600M_img512_TS_training.yaml to set your data paths:

   data:
     dataset_dir: /path/to/man-truckscenes  # TruckScenes dataset location
     radar_maps_dir: "radar_maps/"          # Pre-computed radar maps (relative to dataset_dir)
     bev_conditioning_maps_dir: "bev_maps/" # Pre-computed BEV maps (relative to dataset_dir)

2. Run training

   # 8 GPUs (default)
   bash scripts/train.sh configs/RadarGen_600M_img512_TS_training.yaml
   
   # Custom number of GPUs
   NUM_GPUS=4 bash scripts/train.sh configs/RadarGen_600M_img512_TS_training.yaml
   
   # Single GPU with custom arguments
   python scripts/train.py \
       --config_path configs/RadarGen_600M_img512_TS_training.yaml \
       --train.num_epochs=100

Acknowledgements

We thank the following open-source codebases for their wonderful work: SANA, DC-AE, UniDepth, UFM, Mask2Former.

License

This repository's code inherits its license from the following open-source projects:

• SANA: Apache-2.0
• UniDepth: CC BY-NC 4.0
• Mask2Former: MIT License
• UFM: BSD 3-Clause

The pre-trained RadarGen checkpoint inherits its license from SANA's weights and the dataset it was trained on:

• SANA weights: NSCL v2-custom
• MAN TruckScenes: CC BY-NC-SA 4.0

Please refer to the respective repositories and datasets for full license details.

📖 BibTeX

If you find our work useful, please consider starring ⭐ the repository and citing our paper:

@article{borreda2025radargen,
      title={RadarGen: Automotive Radar Point Cloud Generation from Cameras},
      author={Borreda, Tomer and Ding, Fangqiang and Fidler, Sanja and Huang, Shengyu and Litany, Or},
      journal={arXiv preprint arXiv:2512.17897},
      year={2025}
}