Self-Driving Car Engineer Nanodegree Program
This repository contains the project files for the Behavioral Cloning Project in the Udacity Self-Driving Car Engineer Nanodegree.
In this project, a Convolutional Neural Network (CNN) is designed and trained in order to control a car in simulation. The goal is to record data from a human driver and use the data to train a CNN, which can control the car in a feasible way, hence cloning driving behavior.
The trained neural network is considered successful when it is able to drive the car one full round the track in the given Unity simulation environment.
Different models are trained, validated and tested using Keras and TensorFlow as backend. The models output a steering angle to the self-driving car in the simulation.
In the simulator the car is manually steered around a track for data collection. The image data and steering angles are used for training a neural network and then use this model to drive the car autonomously around the track.
A detailed report about the implementation can be found in REPORT.md
Below you will find a summary of the essential details about the repository, files, installation and required dependencies. The project contains various files essential to the training and validation of the Convolutional Neural Network:
model.py: script used to create and train the modeldrive.py: script to drive the car with the trained modelcnn_models.py: contains CNN architectures for Nvidia and Lenetmodels/model_nvidia_final.h5: trained Keras model based on the Nvidia self-driving car CNN architecturemodels/model_lenet.h5: trained Keras model based on the Lenet CNN architecturevideo.mp4: video of the car driving autonomously around the track for one full roundAdditional videos
This project involves the Term 1 Simulator, which can be downloaded here. The simulator provides measurement data from a three forward facing cameras outputting RGB images of 320x160 px, as well as the steering angle and velocity of the car.
The following dependencies are required in order to run the training.
TensorFlowKeras- Additional Python packages
eventlet Flask h5py imageio Keras matplotlib numpy pandas scikit-image scikit-learn scipy seaborn tornado
Installing additional dependencies
$ pip install eventlet Flask h5py imageio Keras matplotlib numpy pandas scikit-image scikit-learn scipy seaborn tornado pydot graphvizA Singularity defintion file is provided: singularity/singularity.def
The Singularity image was used for training the model on an Nvidia DGX-2.
If you get and error when launching the simulator for macOS Catalina, you may have to chance permission for the simulatior. In a terminal run the following command in the folder where your simulator is located.
$ chmod a+x term1_sim.app/Contents/MacOS/*Simon Bøgh