Project X

This is a modified version of naokishibuya/car-behavioral-cloning

Clone the Udacity simulator from udacity/self-driving-car-sim

Steps for setting up the environment:

• Setup a virtual environment with python 3.8
• Install all the dependencies from req.txt

Training the model

The CNN model we use is the modified version of the model described in a paper published by Nvidia. You can find the paper here. This repository contains a pre-trained model but it may cause limited accuracy, since it was trained with a smaller dataset.

• Run the Udacity simulator in training mode and record your game play to a folder
• This will generate a folder named IMG and a CSV file driving_log.csv. The IMG folder contains all the frames captured from the game play, from cameras placed in 3 different angles ( left, center and right).
• The CSV file driving_log.csv contains a mapping of the set of images captured from a single moment and the corresponding steering angle, throttle, speed and brake values
• The CSV structure will be

  Center	Left	Right	Steering angle	Throttle	Speed	Brake

• Copy the CSV file and the contents of IMG folder into a single folder
• For training the model, run python model.py -d path-of-the-folder-with-training-images-and-csv
• This will print a summary of the model and start training the model.

Running the server with the trained model

For running the server without recording the simulator output :

python drive.py name-of-model-file.h5 

For recording the simulator output:

python drive.py name-of-model-file.h5 run1

where run1 is the name of folder to which the recorded frames are to be stored

You can constuct a video from these frames by:

python video.py run1 --fps 30

This will combine all the frames in the folder run1 in to a video of 30fps. Default fps value is 60.