This is the implementation of SRMRL and this repository is based on garage.
To install locally, you will need to first install MuJoCo. Set LD_LIBRARY_PATH to point to both the MuJoCo binaries (/$HOME/.mujoco/mujoco200/bin) as well as the gpu drivers (something like /usr/lib/nvidia-390, you can find your version by running nvidia-smi).
Clone this repo and construct a virtual environment via pipenv install -r requirements.txt. Then activate the virtual environment with pipenv shell.
The implementation of SRMRL is placed in garage/torch/algos/srmrl.py. Add the garage package into your python path with export PYTHONPATH=.:$PYTHONPATH in SRMRL directory.
To run the experiments, you can use the scripts in script folder with
./script/[filename]
The script assumes you are in the SRMRL directory. Or you can run other tasks using
python example/[filename].py --env_name=[env_name]
The output files including log file and model parameters will be placed in ./data/local/experiment/[EXP_NAME].
The output log file can be visualized with tensorboard.
The file progress.csv contains statistics logged over the course of training.
We recommend viskit for visualizing learning curves: https://github.com/vitchyr/viskit.
After training, use sim_policy.py to visualize the learned policy:
python sim_policy.py --model_dir=[output_dir]
This script generate images and gif file of trajectories.
