✨ CARLA-expert

This repo collects the expert in open-source for CARLA leaderboard. As we all known that the expert is important for end-to-end learning driving. The expert is also the upper bound of learning agent performance, smart and efficient expert sometimes may more important than your agent network.

对于一个端到端驾驶来说，第一步就是收集数据，在真实世界，收集数据时都是以人驾驶车为主；而在仿真世界中 一般采用全局视角写一个专家策略；以此作为收集数据时的策略。此repo主要提供一些已开源 并且在e2e agent上训练使用的一些experts

Here are scores that four experts run in the longest6.xml provided from transfuser branch which is same route length as online leaderboard and with eval_scenario. Click here to open the route file. Check the Evaluation and metrics in the official leaderboard.

It only run one time with 36 routes, so just for reference, pls check results more detailed on data/results.

Method	Driving Score ↑	infra ↑	Route completion ↑
MMFN*	97.00	0.98	98.90
Roach	66.73	0.77	88.61
SEED	77.91	0.90	85.49
AUTO*	63.50	0.65	96.71

*: Please note that MMFN and AUTO have some tricks with god view and actions like force destroy, so that their route completions are higher than others. It's a trick but useful for training data. The origin repo didn't have this, I added here.

Setup

please remember to clone with --recurse-submodules

git clone --recurse-submodules https://github.com/Kin-Zhang/carla-expert
conda create -n py37 python=3.7
pip3 install -r requirements.txt

For people who don't have CARLA [在内地的同学可以打开scripts换一下函数 走镜像下载更快点.. ]

./run/setup_carla.sh
# input version
10.1
# auto download now ...

Set the correct the absolute path and CARLA run path!!!

conda activate py37
# ===> pls remeber to change this one
export CODE_FOLDER=/home/kin/workspace/carla-expert
export CARLA_ROOT=/home/kin/CARLA_0.9.10.1
# ===> pls remeber to change this one
export SCENARIO_RUNNER_ROOT=${CODE_FOLDER}/scenario_runner
export LEADERBOARD_ROOT=${CODE_FOLDER}/leaderboard
export PYTHONPATH="${CARLA_ROOT}/PythonAPI/carla/":"${SCENARIO_RUNNER_ROOT}":"${LEADERBOARD_ROOT}":"${CARLA_ROOT}/PythonAPI/carla/dist/carla-0.9.10-py3.7-linux-x86_64.egg":"${CODE_FOLDER}/team_code":${PYTHONPATH}

Running

🚘 the default one is auto with Kin's modification, please remember to change the path! or open the config to change the path!!

Notice the CARLA will start automatically with this script.

python team_code/run_collect.py carla_sh_path=${CARLA_ROOT}/CarlaUE4.sh absolute_path=${CODE_FOLDER}

A. Roach

This expert used the trained RL for expert. This part of jobs is done by Kait from this repo: https://github.com/Kait0/carla-roach

Questions inside 已知问题：

• This repo slow down the speed that resolve this one. Origin repo: Testing on Town01, finding that the velocity is quite high and the brake distance on traffic light may exceed the traffic light and lead the light may not shown at camera images. 本分支下的速度比较慢 所以不会有这个现象；原repo 在Town01测试时，红绿灯路口由于速度较高，停下时已经超过红绿灯了 导致红绿灯在图片中不可见

B. MMFN

This expert is from https://github.com/Kin-Zhang/mmfn

Questions inside 已知问题：

• because of TTC, the brake will be in advance which the reason may not appear if you are using just one camera for trainning agent. 因为内置的time to collision，有时候刹车会较为提前，如果用一个相机可能刹车原因并不会出现在照片中
• really cautious agent. 有点太谨慎了... 还有一定的修改空间

C. SEED

This expert is from https://github.com/autonomousvision/transfuser/tree/2022/team_code_autopilot

Questions inside 已知问题：

• change lane may have collision according to their new paper. 根据新paper来看，换道情况并未考虑进入
• really slow by default setting, advise to modify the max_speed in their script. 测试的时候 默认配置下速度很慢... 如果介意的话 记改改速度

D. AUTO

This expert if from CARLA self for local planner with walker event involved from mmfn repo by Kin.

Questions inside 已知问题：

• only using the distance, efficient but not smart in long term. 仅以距离为因素进行刹停，有效但是长远来看并不smart

Note

Here are some notes.

1. remove stop infra: Since leaderboard didn't count the stop infraction, it removed in this repo (MMFN and AUTO agent) and submodule. And sometimes the stop sign may also cause some mistake on no reason stop when training the agent
2. for the data, you need collect, pls add by your need, there are some examples inside the MMFN and SEED experts
3. ...

Plan

• Involved the e2e agent also. Big hope... Since different e2e agent need different data form
• Experiemtns all the e2e agent in same expert and test their performance, since in the leaderboard, their agent has train in different expert and different dataset, it should be control the suitation in the experiment. 最好做一次完整的实验，控制所有的变量（数据 专家的一致性） 来评判e2e agent的有效性，从leaderboard看来大家的expert不一致，训练数据大小也各不相同，直接从榜单结果看谁的方法更好并不是最佳的方式
• Write a simple technical report of this one.
• ...

Welcome to contribute!

Acknowledgements

If you use this repo in your research, please consider citing our work where we also discussed the influence between experts performance and e2e agent:

@inproceedings{zhang2022mmfn,
  title={MMFN: Multi-Modal-Fusion-Net for End-to-End Driving},
  author={Zhang, Qingwen and Tang, Mingkai and Geng, Ruoyu and Chen, Feiyi and Xin, Ren and Wang, Lujia},
  booktitle={2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages={8638--8643},
  year={2022},
  organization={IEEE}
}

GitHub code repo:

• zhejz/carla-roach
• autonomousvision/transfuser
• dotchen/WorldOnRails
• Kin-Zhang/mmfn

Here are discussion links on the CARLA leaderboard: https://github.com/Kin-Zhang/carla-expert/discussions

✨✨Stargazers, positive feedback

---