search

Linesight-RL / linesight

AI Plays Trackmania with Reinforcement Learning
https://linesight-rl.github.io/linesight/build/html/
497 stars 42 forks source link

Why not the latest version of trackmania? #8

Closed mileseverett closed 1 year ago

mileseverett commented 1 year ago

As someone with ml experience but not rl experience, why did you go for trackmania NF rather than the latest version? Seems there is a nice interface here: https://github.com/trackmania-rl/tmrl

pb4git commented 1 year ago

While tmrl is an awesome project, it is built as a real-time gym environment. What real-time means, is that the observation happens, then the AI calculates its moves, sends the command while the game engine is still running. This introduces random delays before / after an action. We felt that the difficulty of writing a Trackmania AI was large enough, and did not want to tackle the real-time aspect.

To remove the real-time aspect, we tap into the game engine and pause/unpause every action. This is done is tm_interface_manager.py . The method uses a plugin called TMInterface, which is only available for TMNF, that's why we're not using the latest Trackmania version.

mileseverett commented 1 year ago

Ah got it. Thanks for the explanation