The code for the Automatic Driving System (ADS) used as the test-bed for this experiment can be found in this repository.
Donkeycar S1 platform was used in the experiment. Donkeycar is an open source, easy-to-use, and well-documented Python library that can be used with a self-driving 1:10 scale remote control car.
The scaled autonomous vehicle used in the experiment is shown in the figure below comes pre-assembled with a Raspberry Pi 4, frontal camera, remote-controlled car chassis, battery, and a sensor hat. The frontal camera was the only sensor used in the competition. It was used to collect the data, which means collecting labeled images used to train a CNN supervised classifier autopilot using behavioral cloning.
Donkeycar provides tools to collect and label expert demonstrations, clean inaccurate records from the dataset, and train models using deep neural networks. The Donkeycar autopilot machine learning models are created using a form of supervised learning called behavioral cloning. The behavior of an expert is recorded and associated with labeled data and then used to train the model. The model is the output of the neural network and consists of an array of decision-making algorithms.
The Donkeycar software was built upon open source deep learning libraries such as OpenCV for image processing and Keras for training and running the machine learning models. Keras is a lightweight python neural network library that runs on TensorFlow, which the Donkeycar libraries use to train and run autopilot models. From creating to driving autopilot models, Donkeycar offers a variety of tools that allow fast experimentation.
As shown in the figure below, creating models with Donkeycar involves three steps:
a) To collect data using the car. b) To transfer the data to a host computer for cleaning and training the model. c) Transferring the autopilot model back to the car ready to be tested.
To run the following solution, you must have the Donkeycar library installed both in the vehicle itself and your personal computer or a simple Google Colab repository.
If you would like more information on to create a Donkeycar from scratch you can read my blog post in which I go in detail about how to install the libraries in both the vehicle and your computer, you can always check the Donkeycar official documentation. You can also read my experience in the Delta-X autonomous vehicles competition.
With the Donkeycar installed, run the models available in this repository.
Run the linear model available in this repo and copy it to your local Donkeycar folder on your computer. Make sure you put inside the folder mycar/models. Then to copy it to the Raspberry Pi in your car, use the following command. The rsync command is used to sync the folder models from your computer with the folder models from your Raspberry Pi.
rsync -rv --progress --partial ~/mycar/models/pilot-10k-linear-morning10am.h5 pi@raspberrypi:~/mycar/models/
With the model transferred to the car, you are ready to experiment with the self-driving model.
Make sure you connect to your Raspberry Pi from your computer by connecting via SSH.
ssh pi@raspberrypi
Then connected to the security shell of your Raspberry Pi you can run the following command to run a model.
python manage.py drive --model ~/mycar/models/ptub633-and-attack-08-5k.h5
It is essential to notice that you don't have to train these models added to this repo. They had already been trained. The process of training a neural network involves providing examples of labeled data as input and applying mathematical and statistical concepts such as feed-forward and error backpropagation. Those functions will gradually reduce the error between the predicted output and the desired output across several iterations over the dataset. Each cycle of complete training iterations is called an epoch. The training session is composed of as many epochs needed until the model can no longer improve its prediction performance.