olegjakovlev/linesight

AI Plays Trackmania with Reinforcement Learning

Linesight

Trackmania AI
Linesight documentation

Linesight

Linesight is a reinforcement learning project seeking to push what can be done with AI in Trackmania as far as possible.

Trackmania

Trackmania is a racing game that sacrifices some of the realism of sim-racers for very varied track types with all kinds of tricks like wall riding, stunt jumps and wallbangs. Furthermore, Trackmania was designed for equality of input devices which means that keyboard inputs are a viable way to play and therefore that discrete input algorithms like DQN can be applied. In other words, Trackmania is a deep game which can serve as a benchmark to work on any RL algorithm.

Trackmania Interface

Our work, combined with the efforts of donadigo and Kim of the Trackmania Interface team allow interfacing to Trackmania Nations Forever. Allowing you to programmatically send inputs, get car states, get screenshots, etc... This part of our codebase could be useful to other RL projects.

Results

To our knowledge, Linesight is by far the most advanced AI in Trackmania. It was the first to demonstrate human-level driving around May 2023, with Wirtual playing against it in June. In May 2024, Linesight was the first to showcase beating world records on official campaign tracks.

Now that the project is open-source, can you help make it even stronger?