Rover is a set of environments and tools used to perform the experiments of the doctorate thesis "Deep Reinforcement Learning Multi-Sensor Based Navigation and Control".
This thesis was presented by Luiz Afonso Batalha Marão to the São Carlos School of Engineering, of the University of São Paulo.
The Rover environment's are based in the MujocoEnv class of the Gymnasium package (https://github.com/Farama-Foundation/Gymnasium), and the agents are trained by an adapted version of Stable-Baselines3 (https://github.com/DLR-RM/stable-baselines3) PPO algorithm.
The Rover package was tested in Ubuntu 22.04. Thus, it may not be compatible with former versions.
In your shell, at the desired root folder, run the following commands:
git clone https://github.com/luizmarao/Rover
cd Rover
pip install -e.
This will also download and install every package's dependencies. You may want to do it in a new environment (I use conda).
You may launch your experiments calling run.py in your shell. Here follows an example line:
python -m Rover.run --exp_name=4We_001 --exp_root_folder=~/Experiments --env=Rover4We-v2 --num_timesteps=200000 --n_epochs=50 --num_env=4 --n_steps=4096 --batch_size=8192 --networks_architecture=RovernetClassic --img_red_size='(32, 32)' --conv_layers='[(16, 4, 2), (32, 3, 1), (64, 2, 1)]' --seed=10
The run.py file is configured to perform MuJoCo renderings headless (egl). If you don't have a compatible graphics card, make your own copy of the file and remove this configurations' line.
You may also run experiments using a custom script. However, Rover envs were designed
to run in a SubProcVecEnv wrapper. Thus, they need to be called inside an
if __name__ == "__main__": statement.
To cite this repository in publications:
@misc{rover,
author = {Mar{\~{a}}o, Luiz Afonso Batalha},
title = {Rover},
year = {2023},
publisher = {GitHub},
journal = {GitHub repository},
howpublished = {\url{https://github.com/luizmarao/Rover}},
}
``