Donkey Simulator

The Donkey Gym project is a OpenAI gym wrapper around the Self Driving Sandbox donkey simulator. When building the sim from source, checkout the donkey branch of the sdsandbox project.

The simulator is built on the the Unity game platform, uses their internal physics and graphics, and connects to a donkey Python process to use our trained model to control the simulated Donkey.

My Virtual Donkey

There are many ways to use the simulator, depending on your goals. You can use the simulator to get to know and use the standard Donkeycar drive/train/test cycle by treating it as virtual hardware. You will collect data, drive, and train using the same commands as if you were using a real robot. We will walk through that use-case first.



  • Download and unzip the simulator for your host pc platform from Donkey Gym Release.
  • Place the simulator where you like. For this example it will be ~/projects/DonkeySimLinux. Your dir will have a different name depending on platform.
  • Complete all the steps to install Donkey on your host pc.
  • Setup DonkeyGym:
cd ~/projects
git clone
conda activate donkey
pip install -e donkey_gym
  • You may use an existing ~/mycar donkey application, or begin a new one. Here we will start fresh:
donkey createcar --path ~/mysim
cd ~/mysim
  • Edit your to enable donkey gym simulator wrapper:
DONKEY_SIM_PATH = "/home/tkramer/projects/DonkeySimLinux/donkey_sim.x86_64"
DONKEY_GYM_ENV_NAME = "donkey-generated-track-v0"

Note: your path to the executable will vary depending on platform and user.


You may use all the normal commands to at this point. Such as:

python drive

This should start the simulator and connect to it automatically. By default you will have a web interface to control the donkey. Navigate to http://localhost:8887/drive to see control page.

On Ubuntu Linux only, you may plug in your joystick of choice. If it mounts as /dev/input/js0 then there's a good chance it will work. Modify to indicate your joystick model and use the --js arg to run.

python drive --js

As you drive, this will create a tub of records in your data dir as usual.


You will not need to rsync your data, as it was recorded and resides locally. You can train as usual:

python train --model models/mypilot.h5


You can use the model as usual:

python drive --model models/mypilot.h5

Then navigate to web control page. Set Mode and Pilot to Local Pilot(d). The car should start driving.

Sample Driving Data

Here's some sample drivng data to get you started. Download this and unpack it into your data dir. This should train to a slow but stable driver.