C-Shenron: A Realistic Radar Simulator for End-to-End Autonomous Driving in CARLA

This repository contains the official implementation used for the paper "A Realistic Radar Simulator for End-to-End Autonomous Driving in CARLA". The simulator is designed to generate realistic radar data for autonomous driving tasks, enhancing the capabilities of the CARLA simulator.

Project webpage: C-Shenron project page

Overview

C-Shenron is a high-fidelity radar simulation framework integrated with the CARLA simulator, enabling realistic, physics-based radar data generation using LiDAR and camera inputs. It supports customizable radar sensor setups and generates data suitable for End-to-End (E2E) autonomous driving pipelines, especially for transformer-based models like Transfuser++.

Key features:

• Realistic radar data using physics-based Shenron model
• Supports multiple radar views (Front, Back, Left, Right)
• End-to-end integration with imitation learning pipelines
• Easily scalable data collection and evaluation setup
• Demonstrated performance gains over LiDAR-camera baselines (+3% Driving Score)

Data Collection

First we need to generate bash scripts for both starting carla simulator and data collection. data_generation_bash_scripts.py will generate the scripts into Data_Collection_Scripts directory which has two sub-folders:

1. Start_Carla_Job_Scripts contains scripts to start carla simulator and run the data collection scripts
2. Job_Files contains the data collection scripts

Generating scripts:

python3 data_generation_bash_scripts.py

To start data collection

bash Data_Collection_Scripts/Start_Carla_Job_Scripts/job0.sh

This is an example, you can run any of the files from bash Data_Collection_Scripts/Start_Carla_Job_Scripts.

Refer to parallelization.md in this repository for instructions on how to run data collection scripts in parallel by executing each script in a separate pod.

Downloading the dataset

The dataset can be downloaded from the following link: http://wcsng-41.nrp-nautilus.io:8000/

Training the model

Training:

bash team_code/shell_train.sh

Arguments for team_code/train.py:

1. id - Specifies the sub-directory where the trained model will be stored
2. continue_epoch - Use only when you want to use pre-trained model
   • 0 to train from epoch 0
   • 1 to train from epoch where pre-trained model left it
3. radar_channels - Select radar from carla or simulation
   • 2 to use carla's front and back radar
   • <anything else> to use radar data from SHENRON
4. radar_cat - Select the radar concatenation model from SHENRON
   • 1 to use front and back concatenation
   • 2 to use front, back, left and right concatenation
5. use_radar - To use radar data for training
6. use_lidar - To use lidar data for training

Evaluation

Similar to data collection, we need to generate bash scripts for both starting carla simulator and data collection.

1. Evaluation_Scripts/generate_run_bashs.py generates the bash scripts to start carla simulator and running the evaluation scripts into Start_Carla_Job_Scripts
2. Evaluation_Scripts/evaluation_bash_scripts.py generates the evaluation scripts into Job_Files

You can vary all the evaluation parameters in evaluation_bash_scripts.py.

Running Evaluations:

bash /Evaluation_Scripts/Start_Carla_Job_Scripts/job0.sh

Again, this is an example and you can run any of the files from the Start_Carla_Job_Scripts and parallelize the process by following the above mentioned repository.

Citation

If you use this work in your research, please cite our paper:

@INPROCEEDINGS{11310463,
  author    = {Srivastava, Satyam and Li, Jerry and Mishra, Pushkal and Bansal, Kshitiz and Bharadia, Dinesh},
  booktitle = {2025 IEEE 102nd Vehicular Technology Conference (VTC2025-Fall)},
  title     = {A Realistic Radar Simulator for End-to-End Autonomous Driving in CARLA},
  year      = {2025},
  pages     = {1--6},
  doi       = {10.1109/VTC2025-Fall65116.2025.11310463}
}