This project aims to achieve multi-agent cooperative exploration in sparse landmark complex environments. Landmark complex is an abstract simplicial complex proposed for assisting multi-agent exploration. Although traditional landmark complex exploration methods are frontier-based, our method uses deep reinforcement learning to learn a policy for exploration and a three-stage curriculum to help mitigate reward sparsity. This Github repository includes all necessary codes to reproduce our result in Multi-Agent Exploration of an Unknown Sparse Landmark Complex via Deep Reinforcement Learning.
- Unity 2020.3.35
- ML-agents Release 19
- Python 3.8
- Docker
- First navigate to the Python directory and build the docker image
docker build -t lcserver .- Run the Docker container
docker run -d --name lcserver -p 80:80 lcserver-
Open the root directory with Unity Editor 2020.3.35
-
In the Unity Editor, navigate to Assets/LandmarkComplex/Scenes and open LandmarkComplex_Experiment
-
Press the Play button, and the exploration will begin.
-
By pressing the Play button again, the exploration will be terminated.
-
To get the exploration data, export the container
docker export container_id > res.tar- The experiment data can be found at app/data_output/DRL/out.csv