Skip to content

Latest commit

 

History

History
53 lines (46 loc) · 3.73 KB

README.md

File metadata and controls

53 lines (46 loc) · 3.73 KB
Raw

Torcs Autonomous Vehicle using A3C

AI controller for autonomous cars

Torcs A3C Demonstration Video

The image above is a link to a video that show the latest example of the development.

Todo list

C++

  • [Kallah] Add possibility for c++ to read parameters from .ini file
  • [Kallah] Add more comments for c++.
    • [Kallah] ce903.cpp/h
    • [Kallah] exporter.cpp/h
    • [Kallah] raceengine.cpp/h
    • [Kallah] inireader.cpp/h
    • [Kallah] car.h
  • [Kallah] Decrease the latency in torcs (we implemented a latency on sending the images, dont remember why)
  • [Kallah] Remove default torcs files from the github //(I've started this, but its harder than one would think, file x and y may have the same content but different dates etc, and so the python script I wrote to automatically get rid of anything that was unedited from the default torcs directory was not good enough as it does not ignore that kind of stuff).
  • [Kallah] Allow for change of gears, clutch and brakes and add option to ini parameter
  • [Kallah] Solve memory issue in torcs exporter class. SOLUTION: memory leak due to bad memalloc() with non-sane free() methodology.
  • [Kallah] Allow for multiple torcs clients running by having a port number file increment, I.E. client 1 reads port 200, writes port 201 which the next client uses and so on.
  • [Kallah] look into restart memory leak which is apparently inherrent to TORCS itself. each reload uses approx 100-200mb of ram
  • [Kallah] change the driver instructions to be a ini file rather than a csv file for more readable code, avoid duplicate code and to make the driver instructions more readable.
  • [Kallah] rename the resize_img function in the exporter class to reshape_img, as it reshapes, it does not resize.
  • [Kallah] find and get rid of unused methods in the exporter class.
  • Remove using namespace std in c++ files to remove bad C++ practices.
  • [Kallah] move making images grayscale into c++ for increased performance.
  • [Kallah] move preprocessing steps into c++

Python

  • Tune the parameters for model and tried to make the model consistently learn.
  • Add more comments to python code
  • [Kallah] Clean up the python code (I think some methods are unused and some methods can do with refractoring.)
  • Investigate different model achitectures for the A3C in order to increase performance
  • [Kallah] Remake the communication protocol
  • [Kallah] {DEPRICATED: MOVED TO C++} Make the pre-processing into a class for more readable code
  • Perhaps change to this implementation of A3C: https://github.com/awjuliani/DeepRL-Agents/blob/master/A3C-Doom.ipynb
  • Change to regression rather than classification for steering, maybe add accel too. (should be controlled by .ini param)
  • Allow for change of gears, clutch and brakes to be controlled by A3C, on/off by parameters
  • make the pre-processing image into a boolean image. This could help the model avoid local optima, would be similar to semantic segmentation as seen in https://arxiv.org/pdf/1801.05299.pdf, however without the sementics as onyl the road is maintained in the image fed to the model.
  • [Kallah] Move to grayscale images for faster image pre-processing
  • [Kallah] Rework preprocessing to increase performance

Other

  • Detailed instructions for install
  • Detailed instructions for use
  • Detailed instructions for further development

Installation

To set up and run the project use the files inside the 'build' folder. More instructions can be found there as well.

Credits

original project: https://github.com/dmachlanski/ce903-tesla