C3PIL: Crowd Perception Communication-Based Multi-Agent Path Finding with Imitation Learning

Jing Xie, Yongjun Zhang, Huanhuan Yang, Qianying Ouyang, Fang Dong, Xinyu Guo, Songchang Jin and Dianxi Shi (Under review for RA-L)

Model

We propose a decentralized Multi-Agent Path Finding (MAPF) method C3PIL with Crowd Perception Controlled Communication and generative adversarial Imitation Learning. Our overall model achieves a SOTA level.

Setup

Clone the repository git clone https://github.com/JingX/C3PIL.git and move into the top level directory cd C3PIL.
Create conda environment. conda env create -f environment.yml.
Activate the environment. conda activate C3PIL.

Compile cpp_mstar code

cd into the cd od_mstar3 folder. python3 setup.pybuild_ext.
copy .so object from build/lib.*/ at the root of the od_mstar3 folder.
Check by going back to the root of the git folder, running python3 and "import cpp_mstar".

Train

Generate the expert dataset for imitation learning python il_data.py and save datasets at the data folder.
Train python train.py and save the model at the saved_models folder.

Test

Test the model in environments at the test_set folder.
Run python test.py.

Visualization

Cancel the comment in test.py about make_animation() and run python test.py
See results at the videos folder

JJingXie / C3PIL