We saw the Doom challenge and have been inspired to build an autonomous agent, but we tried to switch the task to autonomous driving as we wanted a computer vision intensive task as well.

What it does

The aim of the project was to build a deep learning agent which is capable of driving autonomously by imitating the gameplay of a human user. The environment chosen for this purpose is the racing game TrackMania Nations Forever. We trained deep learning models for this purpose with Convolutional Neural Networks(CNNs), the best performing one being a 4 layers CNN which achieves an acuracy of 82.5\% on the validation set.

How I built it

In order to run TrackMania and be able to send logical keyboard commands to the game, we developed our framework in a Windows environment. Hence, the major part of our framework is built to be run on Windows, only the model training part being compatible with both Linux and Windows. The framework is built using Python along with libraries for machine learning(Tensorflow, TFLearn), screen capturing and image processing(OpenCV).

Challenges I ran into

We wanted to use reinforcement learning but training an A3C agent proved too complicated for a hackathon.

What's next for Autonomous Racing Agent in TrackMania

Reinforcement learning

Built With

Share this project: