Trackmania Nations Forever gets a Gymnasium Compatible Environment

In this student research-project by the ALR the goal was to investigate on how a reinforcement learning agent can learn in a TMNF : Trackmania Nations Forever Environment.

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.

We are still actively working on this project and would appreciate any hints or bugs that we might have missed.

Thank you for your interest and have fun.

Our experiments show that Trackmania-Gym can train agents that reach at least human-like performance. These agents are also able to generalize to unseen tracks, although with some reduction in performance.

This is one of the agents we traied using Trackmania-Gym, in only around 5 hours on the map simple_validated1. Beware this was not our final agent, much rather an example.

This project is licensed under the GNU General Public License v3.0. See the LICENSE file for details.

Copyright (C) 2025 Simon Wilhelm Schübel
Copyright (C) 2025 Tahmur Hassan Mahmood