This is the code accompanying the paper: "CADRE: A Cascade Deep Reinforcement Learning Framework for Vision-based Autonomous Urban Driving" by Yinuo Zhao, Kun Wu, et. al, published at AAAI 2022.
CADRE is a novel CAscade Deep REinforcement learning framework to achieve model-free vision-based autonomous urban driving on CARLA benchmark. We also provide an environment wrapper for CARLA that is suitable for distributed DRL training.
-
Clone repo
git clone https://github.com/BIT-MCS/Cadre.git cd Cadre -
Create conda virtue environment and install dependent packages
conda create -n cadre python=3.7 conda activate cadre pip install -r requirements.txt -
Download the trained perception model from Google Driver under
carla_perception/ -
Download the Carla 0.9.10 server from official website.
To quickly test the installation, we provide a simple script . To run this example, you need to first start the server and then start the client.
To start server, run the script under scripts/start_server.sh. Make sure to replace the CARLA_ROOT to your own directory.
To start training client, change the command in we provide a training script under scripts/simple_test.sh, you need to change the CARLA_ROOT and CHALLENGE_DIR.
If the installation is successful, then you will see the following two windows on your computer.
To start server, run the script under scripts/start_server.sh. Make sure to replace the CARLA_ROOT to your own directory.
To start training client, we provide a training script under scripts/main.sh, you need to change the CARLA_ROOT and CHALLENGE_DIR. The hyperparamters are configured under /config_files/agent_config.py. We recommend you to change the hyperparameter num_processes to 4 in order to get a more stable policy.
Models and log files are saved under result/
After the training process finished, we recommend you to use the script under scripts/kill_server.sh to kill all servers running on the server.
#!/bin/bash
export CARLA_ROOT=[PATH TO YOUR LOCAL DIRECTORY WITH CaraUE4.sh]
export CHALLENGE_DIR=[PATH TO WHERE]
export PYTHONPATH=$PYTHONPATH:$CARLA_ROOT/PythonAPI/carla
export PYTHONPATH=$PYTHONPATH:$CARLA_ROOT/PythonAPI/carla/dist/carla-0.9.10-py3.7-linux-x86_64.egg # 0.9.10
export PYTHONPATH=$PYTHONPATH:$CHALLENGE_DIR/leaderboard
export PYTHONPATH=$PYTHONPATH:$CHALLENGE_DIR/scenario_runner
export HAS_DISPLAY='0'
python main.py
To evaluate the models, please refer to the scripts under scrips/eval.sh. Please set the pretrained_path and load_episodein eval_cfg under config_files/eval_agent_config.py. It is recommended to use 8 models from different episodes for a more stable policy. You can also change the amount of vehicles/pedestrians and routes in this config files.
After evaluation ends, you can find the evaluation results under ${pretrained_path}/eval/eval_completion_ratio.csv.
This work was supported in part by Shanghai Pujiang Program and the National Research and Development Program of China (No. 2019YQ1700).
If you have any question, please email ynzhao@bit.edu.cn / linda.chao.007@gmail.com.
This project includes some implementations of DANet and the overall evaluation framework follows CARLA secnario runner, carla_project (no license) and leaderboard.
If you are interested in our work, please cite our paper as
@inproceedings{zhao2022cadre,
author = {Zhao, Yinuo and Wu, Kun and Xu, Zhiyuan and Che, Zhengping and Lu, Qi and Tang, Jian and Liu, Chi Harold},
title = {CADRE: A Cascade Deep Reinforcement Learning Framework for Vision-based Autonomous Urban Driving},
booktitle = {Association for the Advancement of Artificial Intelligence (AAAI)},
year = {2022},
}