This repo is a fork of MBRL-Lib, which is a toolbox for facilitating the development of Model-Based Reinforcement Learning algorithms. Our fork adds the needed code to use the Gym-Duckietown environment as a configurable option, further extending the use of the toolbox for autonomous driving tasks.
Before continuing, we suggest you take a look at the readme files for both MBRL-Lib and Gym-Duckietown, as you will need to be familiar with both tools for the following instructions.
We recommend making a python virtual environment for using this project. You can use Anaconda to set up and manage your virtual environments. Use the following command to create a new virtual environment with python version 3.8:
conda create --name RLDucky python=3.8
Then to activate it use:
conda activate RLDucky
You will use this python virtual environment whenever you are running code from this repo. You should use a seperate virtual environment when running code from the duckietown repo, such as in the commands below.
Make a new virtual environment for the Gym-Duckietown repo:
conda create --name GymDucky python=3.8
Then to activate it use:
conda activate GymDucky
The first repo to clone is Gym-Duckietown, and make sure you checkout and use the master branch. Additionally you can install the required python packages for that repo via the command pip install -e . where . specifies the current directory.
conda activate GymDucky
git clone https://github.com/duckietown/gym-duckietown.git
cd gym-duckietown
git checkout master
pip3 install -e .
While trying to use Gym-Duckietown we ran into an issue involving a malfunctioning / deprecated geometry module. If you run into the same problem, you can just comment out that import. So just navigate to the gym-duckietown/gym_duckietown/simulator.py file and comment out the import geometry line.
For the following commands, switch back to your RLDucky environment:
conda activate RLDucky
To use the Duckietown environment seamlessly with MBRL-Lib, we will have to add the gym-duckietown repo as a python module to our python installation. There are two ways of doing this.
Locate your python installation and add a path (.pth) pointing to the gym-duckietown repo.
You can find your python installation (whether it is in a virtual environment or not) by using the Python shell. Use the python command to enter the Python shell, then use the following commands:
python
Python 3.8.12 (default, Oct 12 2021, 13:49:34)
[GCC 7.5.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import sys
>>> for p in sys.path:
... print(p)
...
Which will output something similar to this:
/your/path/to/mbrl-lib
/your/path/to/gym-duckietown
/your/path/to/anaconda3/envs/RLDucky/lib/python38.zip
/your/path/to/anaconda3/envs/RLDucky/lib/python3.8
/your/path/to/anaconda3/envs/RLDucky/lib/python3.8/lib-dynload
/your/path/to/.local/lib/python3.8/site-packages
/your/path/to/anaconda3/envs/RLDucky/lib/python3.8/site-packages
>>>
Then navigate to the /your/path/to/anaconda3/envs/RLDucky/lib/python3.8 folder, where you will find some .pth files. Make a new one called duckietowngym.pth (name is not important, you can call it whatever you would like) and make this the file's contents:
import sys
sys.path.append('/your/path/to/gym-duckietown')
Now restart the terminal and python should be able to find gym-duckietown for imports.
Append the /your/path/to/gym-duckietown to your PYTHONPATH environment variable. You can do this using the command:
export PYTHONPATH="${PYTHONPATH}:/your/path/to/gym-duckietown"
To check that you did it correctly, you can use the command
echo $PYTHONPATH
You will have to do this every time you restart your terminal. If you want to do so automatically, you can add the export line above to the end of your .bashrc file. Note that the .bashrc file is typically a 'hidden file' so you may need to change a setting somewhere to find it using a graphical file explorer.
Clone this repository and install the required python packages:
git clone https://github.com/alik-git/duckietown-mbrl-lib
cd duckietown-mbrl-lib
conda activate RLDucky
pip install -e ".[ducky]"
And test it by running the following from the root folder of the repository
python -m pytest tests/core
python -m pytest tests/algorithms
You may run into issues here if you don't have CUDA installed. If you have a CUDA compatible GPU, you can install it via these instructions. Otherwise you can ignore these tests and simply switch to CPU by replacing "cuda:0" with "cpu" in the /mbrl/examples/conf/main.yaml file.
You can confirm that you have CUDA installed correctly by running the command:
nvcc --version
We will need to use the MuJoCo physics engine for some of our experiments. You can find installation instructions here.
You may also need to modify your LD_LIBRARY_PATH environment variable to get MuJoCo to work properly. To do that you can use the following commands:
export LD_LIBRARY_PATH="${LD_LIBRARY_PATH}:/your/path/to/.mujoco/mujoco210/bin"
export LD_LIBRARY_PATH="${LD_LIBRARY_PATH}:/usr/lib/nvidia"
You will have to do this every time you restart your terminal. If you want to do so automatically, you can add the export line above to the end of your .bashrc file. Note that the .bashrc file is typically a 'hidden file' so you may need to change a setting somewhere to find it using a graphical file explorer.
To confirm that MuJoCo is installed and working correctly, run the following command from the mbrl-lib repo's root folder.
cd /your/path/to/mbrl-lib
python -m pytest tests/mujoco
A few issues with mujoco - and I've encountered this during other projects as well - are about mujoco not neatly attaching to a rendering engine in your software environment. On Ubuntu it seems to try and use the OpenGL rendering engine, and to find it is uses GLEW. But this process doesn't always end up being seamless. Below are a few issues that we ran into:
While trying to run MuJoCo we ran into an error twice relating to something involving undefined symbol: __glewBindBuffer, and the only fix we found (from a Reddit thread) was to install the following packages:
sudo apt install curl git libgl1-mesa-dev libgl1-mesa-glx libglew-dev \
libosmesa6-dev software-properties-common net-tools unzip vim \
virtualenv wget xpra xserver-xorg-dev libglfw3-dev patchelf
This issue is again about the dm_control installation of mujoco, you basically get an error saying mujoco.FatalError: gladLoadGL error, and as described in the aformentioned link, the solution is to find your dm_control package in your python installation, something like:
/home/username/anaconda3/envs/RLDucky/lib/python3.8/site-packages/dm_control/_render/glfw_renderer.py
and modify it to add some extra lines as described here.
We use Weights & Biases for logging and visualizing our run metrics. If you're unfamiliar with Weights & Biases, it is a powerful and convenient library to organize and track ML experiments. You can take look at their quick-start guide and documentation, and you'll have to create an account to be able to view and use the dashboard, you can do so here.
For this project, just specify your wandb username and project name in the main.py file in the following section:
if __name__ == "__main__":
wandb.init(
project="<Your W&B Project Name Here>",
entity="<Your W&B Username Here>"
)
run()To run an experiment you can use commands in the following format:
python -m mbrl.examples.main algorithm=planet dynamics_model=planet overrides=planet_duckietown
Here is another example of a command with shorter episodes:
python -m mbrl.examples.main algorithm=planet dynamics_model=planet overrides=planet_cheetah_run algorithm.test_frequency=2 overrides.sequence_length=10 overrides.batch_size=10
You will see the output of your run in the terminal as well as in a results file created by Hydra located by default at .exp/planet/default/duckietown_gym_env/yyyy.mm.dd/hhmmss; you can change the root directory (./exp) by passing
root_dir=path-to-your-dir, and the experiment sub-folder (default) by
passing experiment=your-name. The logger will also save a file called
model_train.csv with training information for the dynamics model.
To learn more about all the available options, take a look at the provided configuration files.
If you run out of memory, you can decrease the dataset size parameter in the /mbrl/example/conf/algorithm/planet.yaml file. Do not reduce it to anything under 1000 or it might fail.
For those using VSCode (and you really should be if you aren't), you may need a little help with setting up your debugging configuration since MBRL-Lib is designed to be a python module.
If you're new to debugging in VSCode, you can take a look at these instructions.
Click on Run And Debug (Ctrl+Shift+D).
Click on Create a launch.json file.
Replace the contents of the default launch.json file with the following and make sure to change the path to match your installation! The key change here is that we're using the "module" option instead of the "program" option.
{
// Use IntelliSense to learn about possible attributes.
// Hover to view descriptions of existing attributes.
// For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
"version": "0.2.0",
"configurations": [
{
"name": "Python: Current File",
"type": "python",
"request": "launch",
"python": "your/path/to/anaconda3/envs/RLDucky/bin/python",
// "program": "${file}",
"module": "mbrl.examples.main",
"args": [
"algorithm=planet",
"dynamics_model=planet",
"overrides=planet_duckietown"
],
"console": "integratedTerminal"
}
]
}
Our work here is a small addition to the MBRL-Lib project and Gym-Duckietown project. All rights for those projects are held by their respective authors.
mbrl is released under the MIT license. See the mbrl LICENSE file for
additional details about it. See also the MBRL-Lib author's Terms of Use and
Privacy Policy.
See their Summary of Duckietown Software Terms of Use and their Terms and Conditions.