RL environment for large-scale autonomous driving tasks

This repository aims to build a general and extensible RL training and testing environment for the autonomous driving tasks. It is highlighted in the following way:

We incorporate a sophisticated vehicle dynamics and the large-scale traffic and map constructed using SUMO to realize co-simulation of the ego vehicle and its surrounding environment.
We design the basic RL elements in the field of autonomous driving following the interface defined in Gym, such as the state, the action, the reward function, and the reset conditions etc, which can be easily reused or modified depending on your task (see here for an example).
Correspondingly, we develop an analytic model of the environment to facilitate its use in model-based RL algorithms.
We also include a bunch of training and testing environments that we use, for instance, the one for integrated decision and control (see branch master); the one for centralized coordination (see branch centralized_env); ones with specific designed RL elements such as using bird view image as state (cross_nocontrol_fixedgrid_end2end), or using trajectory as action (toyota202003); and ones with different vehicle parameters, or the map (e.g. toyota202012, toyota202012_3lane, toyota202103_exp).

Getting Started

Prerequisites

SUMO >= 1.7.0
Python >= 3.6
Gym >= 1.17.3
Tensorflow >= 2.2

Install

Download this repo on your host.
Register this environment (see here or follow the below steps).
- Create a python package, say user_defined, under your installed gym path /path/to/gym/envs/
- Copying the downloaded files endtoend.py, dynamics_and_models.py, endtoend_env_utils.py, traffic.py, together with the whole directory sumo_files to /path/to/gym/envs/user_defined.
- Add a line in /path/to/gym/envs/user_defined/__init__.py: from gym.envs.user_defined.endtoend import CrossroadEnd2end
- Add the following code in the end of the file /path/to/gym/envs/__init__.py
```
register(
    id='CrossroadEnd2end-v0',
    entry_point='gym.envs.user_defined:CrossroadEnd2end',
    max_episode_steps=200,
)
```

How to use

Use it just as the other embedded gym environments

env = gym.make('CrossroadEnd2end-v0')

Reference

This environment is officially proposed in the paper integrated decision and control and here's a BibTeX entry that you can use to cite it in a publication.

@article{guan2021integrated,
  title={Integrated Decision and Control: Towards Interpretable and Efficient Driving Intelligence},
  author={Guan, Yang and Ren, Yangang and Li, Shengbo Eben and Ma, Haitong and Duan, Jingliang and Cheng, Bo},
  journal={arXiv preprint arXiv:2103.10290},
  year={2021}
}

LJP-luo/env_build