Donkeycar is minimalist and modular self driving library written in Python. It is developed for hobbiests and students with a focus on allowing fast experimentation and easy community contributions.
- Make an RC car drive its self.
- Compete in self driving races like DIY Robocars
- Experiment with different driving methods.
- Add parts and sensors to your car.
- Log sensor data. (images, user inputs, sensor readings)
- Drive your car via a web or game controler.
- Leverage community contributed driving data.
- Use existing hardware CAD designs for upgrades.
After building a Donkey2, here are the steps to start driving.
git clone https://github.com/wroscoe/donkey donkeycar
pip install -e donkeycar
donkey createcar --path ~/d2
python ~/d2/manage.py drive --throttle=0.25
This will start recording data when the throttle is pressed from the controls.
Now you can control your car by going to <ip_address_of_your_pi>:8887/drive
You can use the Web controls or a PS3 controller (joystick).
Run around 20 - 30 laps
The data will be recorded to the ~/d2/data folder.
I find easiest to drive with the throttle maxed. Set a max throttle to a value that allows you drive always at full throttle. This is because the timing of when you take a turn and the angle of the turn will vary with speed. So if your speed fluctuates, your steering will also.
python ~/d2/manage.py train --model=any_model_name --/full/path/to/data/tub_XXX... The tub must be the whole path. You can also load multiple tubs separated by a comma. There cannot be any spaces between each tub path. The model can be any name. This will create a model with the given name in the ~/d2/models folder.
python ~/d2/manage.py drive --model=any_model_name --throttle=0.25This will load the trained model for your drive. When you first start this, it will be in 'User' mode.
You will need to change the mode to 'Local_Angle'. This will cause the steering to be autonomous but the throttle is still controled by the user.
Set the same throttle value as when you trained and max the throttle and let it steer autonomously.
python ~/d2/manage.py train --tensorboard=True --model=any_model_name --tubs=/full/path/to/tub_XX_XX-XX-XX tensorboard --logdir Graph/While training, files will be generated in ~/d2/Graph, these files can get large.
After running this command, a URL will be given to view the Tensorflow process.
Refresh the page to see updates in the training.
Clear the ~/d2/Graph folder when starting a new training.
angle_out_loss, throttle_out_loss and loss are based off the training data.
val_angle_out_loss, val_throttle_out_loss and val_loss are based off the validation training set data.
If you modify the source code, you will need to reinstall the donkeycar. Run these commands to reinstall donkeycar with your modified source code.
pip uninstall donkeycar
cd /path/to/donkeycar/source/code
python setup.py build
pip install -e /path/to/donkeycar/source/codeI created a script file "install_on_desktop.sh" to do all these commands for you.