AI learns to exploit a glitch in Trackmania

I trained an AI in Trackmania with reinforcement learning, and tried to make it learn the hardest technique in this game: the noseboost.

To support my work on Patreon:
•   / yoshtm  

Note: This video was re-uploaded a few hours after the original upluoad due to a copyright issue with the music

Contact
• Discord: yosh_tm
• Twitter:  yoshtm1  
• Mail: [email protected]

Link to download the last map:
https://tmnf.exchange/trackshow/12306053

TAS runs shown in the video:
A01-Race, 17.61 by igntul & Lukalyc:    • [TAS] A01-Race in 17.61 (-6.19)  
A09-Race, 21.08 by gingerroar & plastorex:    • [TAS] A09-Race: 21.08 (-2.05 TAS, -3....  
D12-Speed, 19.94 by threadd:    • [TAS] D12-Speed 19.94 (-18.49 WR / -6...  

Other reinforcement learning projects mentioned in the video:
GT Sophy:    • The Making of Gran Turismo Sophy [Eng...  
OpenAI Five:    • OpenAI Five: Dota Gameplay  
RLGym: https://www.twitch.tv/rlgym?lang=fr
Linesight:    / @linesight-rl  

Thanks to Donadigo, for TMInterface!
https://donadigo.com/tminterface/

The setup used in all maps to put the car on its nose was inspired by this map:
https://tmnf.exchange/trackshow/9042183

On a more general note, this video would not have been possible without the passion and creativity of the TAS community, who discovered the noseboost in 2021. Special thanks to them :)
More information on Trackmania TAS here:   / discord  

You can find a list of the musics I used at the end of the video.