I suspect that this is a very difficult problem for an AI to solve. It requires advanced planning and game knowledge that I don't think that current AI paradigms are capable of.
Videos:
https://www.youtube.com/watch?v=yzqq8I3JFz4
https://www.youtube.com/watch?v=9w_PrEOeMjw
Spreadsheet of best TAS times:
https://docs.google.com/spreadsheets/d/1R1N4mQkRyQemJgDAGh54v9_n4fLW4Z3EGp71xrIBRIY/edit#gid=0
This question will resolve true if an AI is able to beat or equal 50% of human-created TAS across the TMUF/TMNF Nadeo campaign maps at any point before December 31st 2025.
Related questions
Not TAS, but multiple people seem to be working on something adjacent here for competing with human real-time runs?
https://www.youtube.com/watch?v=kojH8a7BW04
https://www.youtube.com/watch?v=cUojVsCJ51I
Could it be done ... definitely!
Will it be done ... depends if some one invests the time to do it!
Even without AI, computers were able to beat chess Grandmasters just by using their sheer processing power. (AI chess programs have since been developed such as AlphaZero)
Chess has very few rules and a "pause button" between rounds which fits in perfectly with simulating future rounds and planning ahead.
Now the real question becomes "How would you use AI to beat human performance in Trackmania?"
You could in theory create an AI that was able to do the runs in real time a supposed SuperPlayer.
Another way would be to train the AI on the layout of each track or at least the desired track, quirky game physics, game mechanics and other factors and have it create a TAS input list.
With AI the goal is to arrive at those answers faster than the brute force method mentioned above and with less energy used.
Seems unlikely without some significant guidance, given how much these ~minute long runs hinge on taking exactly the right sequence of sub-second actions at the right moment. An RL agent isn't going to stumble it's way into this. Maybe with some heavy behavioral cloning to get it started off? But I think the glitches it needs to exploit are just too precise and require too much chaining.
