Nvidia created a multi-layer CNN called PilotNet, in 2017, that was able to steer a vehicle by just showing it a series of images or video. This was a compelling demonstration of the power of neural networks, and in particular the power of convolution. A diagram showing the neural architecture of PilotNet is shown here:
PilotNet neural architecture
The diagram shows the input of the network moving up from the bottom where the results of a single input image output to a single neuron represent the steering direction. Since this is such a great example, several individuals have posted blog posts showing an example of PilotNet, and some actually work. We will examine the code from one of these blog posts to see how a similar architecture is constructed with Keras. Next is an image from the original PilotNet blog, showing a few of the types of images...
