- Видео 9
- Просмотров 1 212 722
Linesight
Добавлен 4 май 2023
We train an Artificial Intelligence with Reinforcement Learning to play the game Trackmania Nations Forever, and post videos showcasing the progressive improvement of our A.I.
This channel is a collaboration between pb4 (github.com/pb4git) and Agade (github.com/Agade09).
You can contact us at the address pb4videos (at) gmail.com, via our github, or on Discord (server: discord.gg/tD4rarRYpj and channel: discord.com/channels/847108820479770686/1150816026028675133).
This channel is a collaboration between pb4 (github.com/pb4git) and Agade (github.com/Agade09).
You can contact us at the address pb4videos (at) gmail.com, via our github, or on Discord (server: discord.gg/tD4rarRYpj and channel: discord.com/channels/847108820479770686/1150816026028675133).
I Trained an AI for 2 Years on Trackmania. It's Breaking Records.
I trained an AI that plays Trackmania with reinforcement learning. It's getting good enough to challenge Official World Records. This is the story of how my AI reached such an incredible level.
Link to raw run playlist: ruclips.net/p/PLZ-JKrjYnD1GBlp-WS3Mr8bKmqsqSSO-U
Link to replay pack: drive.google.com/file/d/19p10V3g4RfFUoUUdVWNN2RXwJmcXfHC2/view?usp=drive_link
Link to input files: drive.google.com/file/d/12Jfp6nDNhAb7vQIeUIao_VU0W8652s_r/view?usp=drive_link
Project Github: github.com/Linesight-RL/linesight
Game: Trackmania Nations Forever
Buymeacoffee: buymeacoffee.com/linesight
00:00 Intro
00:43 map5 (v1 and v2)
03:58 A01
05:10 map5 (v3)
07:27 Official Campaign intro
08:42 A02
11:11 D06
14:22 Off...
Link to raw run playlist: ruclips.net/p/PLZ-JKrjYnD1GBlp-WS3Mr8bKmqsqSSO-U
Link to replay pack: drive.google.com/file/d/19p10V3g4RfFUoUUdVWNN2RXwJmcXfHC2/view?usp=drive_link
Link to input files: drive.google.com/file/d/12Jfp6nDNhAb7vQIeUIao_VU0W8652s_r/view?usp=drive_link
Project Github: github.com/Linesight-RL/linesight
Game: Trackmania Nations Forever
Buymeacoffee: buymeacoffee.com/linesight
00:00 Intro
00:43 map5 (v1 and v2)
03:58 A01
05:10 map5 (v3)
07:27 Official Campaign intro
08:42 A02
11:11 D06
14:22 Off...
Просмотров: 843 398
Видео
Trackmania AI Learns To Drift and Beat Pros ? | Hockolicious
Просмотров 348 тыс.10 месяцев назад
We trained an AI to play Trackmania on Hockolicious, one of the game's most prestigious tracks. The AI set a strong time, beating the 2012 World Record previously set by CarlJr. Link to map: tmnf.exchange/trackshow/414041 Link to replay: drive.google.com/file/d/1b0p2QSXAea1v7lVCGuWkNr-5xQ6PdqHo/view?usp=drive_link Game: Trackmania Nations Forever Project Github: github.com/Linesight-RL/linesigh...
Superhuman Trackmania AI Demo | map5
Просмотров 28 тыс.11 месяцев назад
Linesight project. We used reinforcement learning to train an AI that plays Trackmania. On its training track, the AI is faster than the current human world record. We put this video out as a challenge: can a human still overcome our AI? Game: Trackmania Nations Forever (TMNF) Map: tmnf.exchange/trackshow/10460245 Replay (.gbx file): drive.google.com/file/d/1jfOyhRQCvOC5XSEhhajE2OvmJcCHKDjR Wir...
AI Plays Trackmania - Bloopers
Просмотров 3,3 тыс.Год назад
In this video, and AI is trained with reinforcement learning to accumulate speed and finish a map as fast as possible. The AI learned a behavior where it turns around right before the finish line. This is not a one-off mistake, the AI repeatedly did similar things in back-to-back runs. Can you guess why ?
AI Plays Trackmania - Map5 2:04:91
Просмотров 9 тыс.Год назад
The AI is trained via reinforcement learning. Game: Trackmania Nations Forever (TMNF) Map: tmnf.exchange/trackshow/10460245 Replay (.gbx file): drive.google.com/file/d/1hp1Mz0ooR2YBNpqNjvvxjFrBOGXs8DYD/view?usp=sharing
AI Plays Trackmania - Training Progression Side by Side
Просмотров 3,2 тыс.Год назад
In this video, an AI is trained via reinforcement learning. In order from the top left corner, top right corner, bottom left corner and bottom right corner the AI has received progressively less training time. The video compares the lines taken by the various AIs in different parts of the map. Game: Trackmania Nations Forever (TMNF) Map: tmnf.exchange/trackshow/10460245
AI Plays Trackmania - Map5 2:07:00
Просмотров 1,1 тыс.Год назад
The AI is trained via reinforcement learning. Game: Trackmania Nations Forever (TMNF) Map: tmnf.exchange/trackshow/10460245
AI plays Trackmania - Map5 2:09:12
Просмотров 1,1 тыс.Год назад
The AI is trained via reinforcement learning. Game: Trackmania Nations Forever (TMNF) Map: tmnf.exchange/trackshow/10460245
(Teaser 01) AI learns to play Trackmania with reinforcement learning
Просмотров 2 тыс.Год назад
This video is a recording of an AI currently training to finish a custom map as fast as possible with reinforcement learning. This run was played on Trackmania Nations Forever, with TMInterface to link the game and our AI. The run was played on a custom map. The .Gbx map file is available for download here : tmnf.exchange/trackshow/10460245 The AI took 2mn12s25 to complete the map during that t...
Very good video, amazing use of ia power
what's way more unbelievable is the fact Riolu still hasn't said I'm sorry for cheating. Dude literally banished himself from the internet
This is an incredible project, good job, and thank you for sharing
I have a couple of thoughts. First off, I get the impression you trained the AI on a case-by-case basis for each track. This seems a bit like cheating, as opposed to making a more general one that works well across tracks. The next thoughts are more technical. Instead of downscaling and making it greyscale, you could instead keep a somewhat higher res but run edge detection and reduce to a smaller bit depth as well as greyscale. Just an idea. Finally, on the tracks that do not have as much wall motion, you could cut off the top to save training time and reduce model complexity.
Very nice code, GG!
This was super interesting and entertaining and I don’t even play TM. Thanks.
I have been watching this journey since you started posting these. I remember the first try, the model couldn't even finish the map. It's amazing to see the progress in the model and also how you have adapted to training the model. This is really freaking awesome.
It isn't leaning to drive, it learning the track. A true AI would bee able to take the experience from the track and implement it onto another without having to start from scratch each time.
Interesting video. I hope AI Speedrunning is gonna become a thing lol
6:30 if the fitness algorithm checks car speed and the car alignment compared to the track it is easy to find it )IMO), but if u program the fitness algorithm to punish touching walls heavily it could still dump those.
Have it climb the Deep tower!
pretty sure someone already trained an AI that does keyboard input and also with a lot more gpu horsepower.
That's just a bot running an efficient program. If there was AI, you would not be needed to train it and using BOTS to break records is CHEATING. Please stop or run these cheats on private lobbies if at all allowed by the ToS.
I always want to learn how to make Ai but i never find a good place to it
24:20 I can hear the smirk lmao
Incredible dedication, lovely presentation!
It's a "child" dream of mine to put AI on all those competitive multi-player RTS and watch it figure out new strategies or optimize the meta just like Alpha Star. But I don't know how to do it.
That 20fps seemed to have really helped me understand the surroundings there. I think i could do it
Let it try the deep dip map
Incredible achievement. Two days on a single map to crush most humans, and it's still a work in progress? Absolutely wild.
I guess tokenizing the image will help by bringing "infinite" resolution where it is needed. Or at least try to encode-decode your screenshots and let your AI use the much smaller encoded data.
we all know yosh did it better and frankly this gives off copy vibes
the Ai IS going to play gamea and make art while we toil in the field to make enough food to eat...
What if you made the intended route have punishments instead of rewards and gave rewards only for reaching the end of the map, and additionally gave multipliers to the rewards the quicker it goes (EX: 0.01 = 1 reward) and see it try to figure the next most optimal route through brute force. Possible to make it find different short cuts through trial and error. Also since it uses screenshots to see the track, could you use videos in the same quality and fps to let it see how the track is most optimal? Could it learn from researching and watching videos?
i get in the end its a victory for ai on the first win. but seems unfair sense humans already beat the ai twice. but its a great video none the less. really fun seeing the process
can't beat the bots
I agree that it makes it a little too easy for the AI if it already knows a path to follow and make small improvements on vs having to discover shortcuts or tricks like the fake path in the minitrial
As a developer, I tried wetting my beak in AI programming and holy *&#% I almost fried my brain. 💀 Some people just have the gift and some like me don't hehe 😂
Been following your journey, it's absolutely mesmerizing work, very impressive and inspiring :)
AI doesn't exist, there is no such thing as artificial "intelligence". These programs are not "intelligent", they are not sentient, they are just following instructions. A more accurate description would be "machine evolution".
They should have just had you code the AI F1 cars at Abu Dhabi
Not really that impressive, if you think about it, players do the same thing, run the track over and over again until they get their best lines, AI is just able to do simulations non stop.. The AI isn't doing it smarter, it's just doing it more times in a more methodical manner. This AI is as smart as a password brute force program where it just tries all possible combinations
Can it climb deep dip?
Using multiple agents Smiths from Matrix Reloaded during showing of multiple cars training would be perfect
This is a pretty good showcase of AI’s strengths and weaknesses. AI is good at high performance requirement actions that have lots of data to be trained on since it’s technically perfect. However, it’s not good at novel problems (non-intended shortcuts and the one challenge map that requires you to go through a tiny hole to stay “on track” for example). While this is a very cool project, I unfortunately don’t see it being useful for things like scouting or shortcut hunting.
So you're saying that I can use a fruit fly to get Trackmania World records?
does it generalize? can a car that learned to drive on track A, be good in track B?
Been wallbanging since F1 Race on Gameboy. Used it later to win the endurance races in GT's first track... hands free.
19:50 omg you just gave them anxiety lol
Is there one AI per map? Difference with the OpenAI and Google are that they can handle a new map based on earlier training
Now what about Grand tourismo
I know this was edited in da Vinci resolve just from the title screen animations, cuz I have used it before 😁
trackmAInia It always has been there
I was just wondering right before you mentioned TAS if these kind of inputs can really even be compared to human runs. Its amazing to watch, the process, the little history lessons, and the organization of your experiments. I don't even play Trackmania, but it's fun to watch a visual process of improvement, with numbers and objective comparisons.
I like how you can see some of the AI still take the wrong path on the minitrial map starting at 20:41 i wonder what the other AI's think 😅
Great video, loved the matrix motifs
TAS vs AI would be insanely interesting
linesight 🔛🔝
damn, fruit fly got hands
Fore me it is always inferior. You put skilled human on new track and he will use its skills to master it in 10 runs. You put master AI of 1 track on new track.... and it needs 1 year and millions of runs to even compare to human.