Behavior Cloning

The goals / steps of this project are the following:

• Use the simulator to collect data of good driving behavior

• Build a convolution neural network in Keras that predicts steering angles from images

• Train and validate the model with a training and validation set

• Test that the model successfully drives around track one without leaving the road

• Summarize the results with a written report

Data Collection and Visualization

I use the simulator to collect data of good driving behavior. First, I collect the data of five laps' correct driving. Then I collect the data of one lap's reverse driving the generalize the data. Finally, I collect one lap's data of sine driving so that the car can learn to go back to the center when it is off the road.

I do cropping on the image and here is the result:

Here is the distribution of data:

Model for training

I first tried Lenet for training a model. However, after times of trying, I found that the performance was not good. So, finding on the Internet, I use the NVIDIA's model structure. Since it has five convolutional layers, it can capture many details, and since NVIDIA uses this model to train real car, I think it definitely has the ability to train my car. Based on that, I changed it a little bit by adding Dropout layers to prevent the overfitting. Here is the visualization of my model structure. By the way, I use the Adam optimizer.

Training Set and Validation Set

I use the generator function provided on Udacity's course. I also do a little change on it to augment data. I randomly choose the camera image. If it is the right camera, the angel will minus -0.25, if it is the left camera, the angle will add 0.25. Also, I flip the image horizontally to generalize the data.

Here is the loss:

I also tried for 10 epochs, but the performance is worse, so I chose to train for only five epochs.

Test Video

See "behavior_cloning.mp4"