Installation

This page guides you through the installation process to get up and running with DRLoco. It should take about an hour, but might also get tricky depending on your system and require some googling to solve unexpected issues. So be prepared and best luck!

Warning

This repository was developed and tested on Ubuntu 16.04 and Ubuntu 18.04 using Python 3.7. The installation instructions are the same for both Ubuntu versions and should also apply to newer Ubuntu versions.

  1. Install Anaconda following the official installation guide.

    1. Anaconda, or conda for short, is a data science toolkit and package manager that allows you to easily install open-source python packages and organize them in separate environments.

  2. Create a new conda environment with Python 3.7 and activate it (you can change the environment name drloco to evry other name you like)

    conda create -n drloco python=3.7
conda activate drloco

  3. Install PyTorch using the official installation guide choosing the following options:

    ../_images/pytorch_installation_configuration.png

  4. Install MuJoCo and mujoco-py following these instructions.

  5. Install Stable-Baselines 3 with the command below

    pip install stable-baselines3[extra]

    1. If sth. goes wrong, remove the ‘[extra]’. This will leave out some packages, that you might need to install later separately, e.g. tensorboard.

    2. If there are still issues, please check the official installation instructions of Stable Baselines.

  6. Install Weights & Biases for beautiful and professional logging of training results and seaborn for plotting:

    conda install seaborn
conda install -c conda-forge wandb

  7. Clone the DRLoco repository to your local PC (Link to repository).

    1. In the following, we refer to the absolute path of the local folder on your PC you cloned the repository to as /local/path/to/repo/.

  8. [OPTIONAL] If you want to distinguish between running your code on a weak local laptop and running it on a powerful remote server, adapt is_remote() in scripts/common/utils.py. Please refer to this function’s doc-string for more details.

  9. To check if installation was successful, run the script scripts/run.py. To run the script from command line, use:

    cd /local/path/to/repo/
python scripts/run.py

    1. If everything went well, a simulator window should appear. Press [Space] to start the simulation and you should see a MuJoCo model walking in circles for 10 seconds.

    2. If you get import errors, install the the missing packages in your conda environment.

    3. When you get the error GLEW initalization error: Missing GL version, please add the following text at the end of the file ~/.bashrc:

      export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libGLEW.so:/usr/lib/x86_64-linux-gnu/libGL.so

      1. If that still not solves your issue, please refer to this bug report.

Installing Additional Features

Warning

Video recording might be broken in the moment. To see the performance of an agent, please run your model using run.py. Please read Running an agent for more details.

  1. To record videos of the generated walking gait at the training’s end, we need to further install ffmpeg:

    conda install -c conda-forge ffmpeg

  2. If you want to record the video on a remote server without a UI, in addtion install xvfb with sudo apt install xvfb. To execute the training script run xvfb-run python /path/to/train.py.