Welcome to the Documentation of Trackmania-Gym

Trackmania-Gym is a Gymnasium Compatible Reinforcement Learning Framework for Trackmania Nations Forever, designed for real-time racing, agent analysis and autonomous driving research.

This framework offers a modular design for experimenting with observations, rewards and terminations, allowing you to focus on actually designing your agent, and not getting lost in annoying implementation-details.

How to read

If installation is complete, continue with 1. Reference Line - outlaying the fundamentals of our measure of progress along the track 2. The Structure of the project - undersand how the framework is organized and get a feel for the environment 3. The Managers - as the basic building blocks of the environment 4. Our testing framework - essential for live visualization of the output of the managers

Config-Based Launch

Trackmania-Gym uses Hydra to manage the configuration of our project. The base-configuration is configs/train.yaml. (Almost) all scripts in scripts/ automatically laods theis configuration, which includes all sub-configurations (environment, agent, logging, etc).

Get familiar with the config-setup, by skimming through train.yaml and following the paths to other configuration-files, such as rl_env/single_agent_env.yaml.

Hydra-Instanciation

In order to not hardcode values into our code, Trackmania-Gym uses Hydra-instanciation to instanciate certain classes, like for example the ObservationManager. Every implemented component has its own configuration file, for example each implemented ObservationManager variant corresponds to a YAML file that Hydra uses to build the class at runtime.