Autonomous Remote Controlled Car

Objective

The goal of this project was to build an autonomous remote controlled car using deep learning. The car is embedded with a Raspberry Pi and a Convolutional Neural Network (CNN) to predict in real-time the speed and the steering angle. We used an ultrasonic sensor to stop the car when encountering obstacles and another CNN for detecting stop signs.

Championships

We participated to two competition. We got 3rd at Iron Car Summer 2018 and 1st at RobotCars Winter 2018. Check out our best performance here.

Version 1.0: SunFounder

We used the SunFounder PiCar-S Kit V2.0 for Raspberry Pi with Raspberry Pi 4B and TF card for the hardware. It costs around 180$ and can be built in an hour or two. The final presentation of this project can be found as a pdf or a powerpoint file. You can find some videos of the car's performance here.

Version 2.0: DonkeyCar

For the second version of this project we used the DonkeyCar kit to build a faster and steadier car model. We generated some data using PIL by drawing lane lines on floor pictures. You can find the code of the data generator in smartcar\simulator. The version 2.0 is much better than the first version as you can see in this video. Our university EPITA highlighted our project in this article.

Installation

Run the following line to create the appropriate conda environment:

conda env create -f environment.yml

Caution

This repository aims at giving a sample project for building a self driving rc car. The logic for the client-server is only working with the hardware used for this experiment. We highly advise you to understand the smartcar\learn module for knowing how to train the model and the smartcar\server\auto_drive.py file for the autonomous driving logic.

Generate your own data

You can use our script to generate your own data. Get many pictures of the floor, put them in a directory and run the following command:

python scripts/generate.py -i image_dir -n 10 -p output_dir

where -i option is for specifying the input data, -p the output directory and -n the number of images to generate.

Train your own model

If you have your own data i.e (image, steering_angle, speed), we created a script that allows you to train you own model. In a given directory, the images should a the form unique_id.jpg and the label as the json file unique_id.jpg. The json file should have the following format:

{
  "angle" : 0.0,
  "velocity": 0.0
}

Then cd at the root of the project and run the training script:

python scripts/train_pilot.py --help

You will get the following output:

  -h, --help            show this help message and exit
  --data_dir [DATA_DIR]
                        The directory containing the training data.
  --out_dir [OUT_DIR]   The output directory to store best model and training
                        curves.
  --lr [LR]             The learning rate.
  --batch_size [BATCH_SIZE]
                        The batch size.
  --epochs [EPOCHS]     The number of epochs.

Fill these arguments according to your configuration, here is an examples:

python scripts/train_pilot.py --data_dir=data/tracks --out_dir=tmp --lr=1e-4 --batch_size=128 --epochs=20