PGDrive: An Open-ended Driving Simulator
with Infinite Scenes from Procedural Generation

Quanyi Li2*,  Zhenghao Peng1*,  Qihang Zhang2,3,  Cong Qiu2,  Chunxiao Liu2,  Bolei Zhou1 
1The Chinese University of Hong Kong, 2SenseTime Research, 3Zhejiang University
Code | Documentation | Paper
[EN]  [中文]
Overview of PGDrive Simulator
To better evaluate and improve the generalization of learning-based driving systems, we introduce an open-ended and highly configurable driving simulator called PGDrive. PGDrive can generate a diverse set of driving scenes through procedural generalization from basic traffic building blocks. Currently the simulator is used to study the generalization of the driving agents trained from reinforcement learning. See paper for more detail.
Procedural Generation of Driving Scenes

We first define the elementary road blocks as follows,
we then follow the proposed algorithm of procedural generation to synthesize maps:
We exhibit more generated maps as follows, which are further turned into interactive environments for reinforcement learning of end-to-end driving.

Result of Improved Generalization

We show that when trained with more procedurally generated maps, the driving agents from reinofrcement learning have better generalization performance on unseen test maps, and can handle more complex scenarios. The detailed experimental results are in the paper. You can reproduce the experiment through our generalization experiment code.

The demo video of the generalizable agent is shown as follows. You can run the agent on your local machine through the provided example in the simulator codebase.

Download the video.

Citation
If you find this work useful in your project, please consider to cite it through: 
@article{li2020improving,
  title={Improving the Generalization of End-to-End Driving through Procedural Generation},
  author={Li, Quanyi and Peng, Zhenghao and Zhang, Qihang and Qiu, Cong and Liu, Chunxiao and Zhou, Bolei},
  journal={arXiv preprint arXiv:2012.13681},
  year={2020}
}