GTAVCNNProject

Goal: Use deep learning related algorithms to autonomously drive a car in GTA V environment in real time

Questions:

  • whether should we solve it using end-to-end driving or other techniques?
  • where are the resources for sending command to GTA V using deep learning in Windows?
  • how to collect datasets in GTA V?

Reconsider the autonomous driving simulators:

Simulator's pros and cons:

1. GTA V

pros:
* has near-real environment with red lights, stop signs, pedestrians and so on
* may be build in OpenAI's universe environment in future
cons:
* not an open source environments. Therefore we may have to concern about the license issue 
* by now, no information can be got about the real positions of other vehicles and pedestrians in the game 
* have to be run in windows environment
* I doubt one GPU can run both GTA V and deep learning (ConvNet) in real time (maybe even 10 FPS). In other words, high hardware requirements

2. TORCS

pros:
* can get real time data of other car's real locations
* have existing source codes for self driving in TORCS
* low hardware requirements 
cons:
* no pedestrians, lights and stops signs available

Online resources:

1. Deep drive(include source code)

2. "A plugin for GTAV that transforms it into a vision-based self-driving car research environment."