Self-driving car that learns from human drivers - part 1

** TORCS is an open-source freeware car racing simulator that is available for Windows, Linux, Mac OS X, Amiga, FreeBSD, MorphOS, and AROS. Visit http://torcs.sourceforge.net/ for more details.

** If you want to know easy way to install TORCS on Ubuntu, visit 2017/07/18 - [TORCS] - How to install TORCS on ubuntu 16.04LTS.

I am on a project "teaching a car how to drive". As a preliminary work, I am trying to teach cars in TORCS. Lots of driving experience data (for example, speed, steering wheel angle, position, etc) is needed in general if someone wants to teach a car how to drive successfully. Obtaining (or collecting) the data of huge size from a racing car controlled by a human is non-trivial. 

So I decide to build a driving policy algorithm for data collection. The nineteen cars in the following video drive themselves according to the algorithm, which includes Anti-collision breaking system, Anti-collision steering wheel control system etc. 

 

<sceenshot>

<video>

 

The left part of the video shows the racing cars. The right part of the video show how the host racing car (red box) recognizes the opponent cars (black boxes) in the track. The large blue box centered at the host car indicates the region where the host car cares about for lane change. So the color of the opponent cars in this area become blue (from black). The yellow box indicates the opponent car that the host car would like to takeover. The host car waits until there exists a room in the blue box region for take-over if some opponent cars occupy the region.
 

Note that this is not the final version. So some cars sometimes bump against each.