AI Learns to play Geometry Dash || Part 2

Let's make an AI to play Geometry Dash!
In my last video I coded a simple clone of Geometry Dash and in this video, I make an AI to play it using neural networks. The neural networks evolve based on which players make it the furthest in the level, and by the end they truly are unstoppable! I hope you guys enjoy the video and make sure to subscribe if you liked it!
LINKS
▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀
Support the channel: ko-fi.com/codenoodles
Itch.io: codenoodles.itch.io/
GitHub: github.com/OfficialCodeNoodles
█▀ █ █ █▄▄ █▀ █▀▀ █▀█ █ █▄▄ █▀▀
▄█ █▄█ █▄█ ▄█ █▄▄ █▀▄ █ █▄█ ██▄

[@CodeNoodles]

Hey, I decided to make a part 3 to play some more levels as you guys requested! kzbin.info/www/bejne/d2Ste6agj7GFidE

[@JavaCake]

:)

[@laty9f1]

Hey man I love your video but please make the screen behind your png thing dark instead of white. I got my retinas burned out at 2 am😅

[@TrideGD]

This AI is WAY too human-like, it struggled with the triple spike. The AI takeover is upon us

[@DeMooniC]

Yo what's up tride lol And yeah I didn't expect it at all to get so far just in the third attempt and struggle with the triple like a real human would do xd. Pretty insane stuff

[@UnknownedYt]

Tride

[@8rbit]

i didnt excpect tride to be here lol

[@HEJOK254]

It even learned straight fly

[@dirhi]

hi hello hi

[@abugidaiguess]

the AI struggling at 63% for generations is pretty relatable tbh

[@anonymous_poketuber2300]

Tru

[@Herobrinegrn1]

How isn't it relatable?

[@freeeng01]

Haha, yes

[@MathHunter]

love how 1 icon struck the whole ship thing

[@Dandailo]

Its just a triple spike tho

[@BusterBrown1217]

The network literally found the MOST anxiety inducing way to fly through that part. I'm not sure if I like it or not.

[@mazenyasser7208]

Because it only clicks when it's absolutely necessary, not like us who may click for the fun of it.

[@killcrap1]

I was stressed watching the first time it got to 63 percent, it had so many near misses

[@tungleson7066]

​@@mazenyasser7208Not a player but aren't that the best way of playing?

[@circumplex9552]

​@@tungleson7066no, its much more risky to do minimum clicks in ship sections

[@AceTechn0]

Dang Generation 3 carried hard, that last cube seriously changed the Best Attempt from like 12% to 63% solo

[@ObvSilent]

lol

[@pandaframe3985]

Even AI can fluke

[@darkfir3.]

@@pandaframe3985 fr

[@oh-no-not-me]

Fr 😭

[@meesvanhaaren4459]

Ofcourse it died at the tripple spikr

[@Commander_Chopper]

It would be very interesting to see how the winner network would perform on a completely different level.

[@firefoxgamingpg3d]

Make it perform on slaughterhouse or another extreme demon let's see how It performs

[@_Fuscous]

@@firefoxgamingpg3d He would have to rebuild the entire level in his remake, meaning he would need to implement every single feature that is in slaughterhouse and not in stereo madness from scratch. I think he's going to stick with 1.0 levels.

[@niek_de_piek]

@@firefoxgamingpg3d Remember the neural network only works with Cube and Ship at the moment, because he used an input node for isFlying, instead of one for every gamemode.

[@thatoneguy9582]

@@_Fuscous hard machine or some shit then

[@Md5offical]

@@thatoneguy9582 back on track or poltergeist maybe

[@puzzLEGO]

Now that the ai is trained, I wonder how it would perform on harder and harder levels

[@Cynderace]

does it pain anyone else to see every single one of the artificial intelligence icons die to the same triple spike at 63% over and over until gen 11

[@freeeng01]

Actually, it behaves as my real neural network, when I started to play this game :D

[@NerdyRulerz]

I love how the nature of the AI means that it does the flight sections in a way that looks as half-assed as possible while still surviving

[@ironici]

I love watching the nodes and connections change color. Since the input requirements are so simple, you can actually make sense of what the AI is thinking (unlike complicated networks)

[@Orestekoa]

It would be nice to see the AI learn some harder levels

[@absoluterightsir4574]

Yeah, like Slaughterhouse or Bloodbath

[@OverHeatVD]

@@absoluterightsir4574 That would probably be too extreme. *no pun intended of course*

[@BritishBoy]

Hey! Thanks for the part 2! I was super hyped for this 😁 Also thanks for featuring my comment in the video :O the game is looking a helluva lot better now too! Hats off to you good sir ;)

[@CodeNoodles]

Thanks, and thanks for the suggestion!

[@BritishBoy]

@@CodeNoodles anytime ;)

[@evandobbs6030]

I live that you explained the why behind some of the code in the neural network part. Most people just make it and implement it without explaining to the viewer what it does, and why.

[@StormSB23]

You could probably make a script that runs while geometry dash is running, and have it control the mouse or space bar so you can train it on the other mains levels, and even custom player levels.

[@mollthecoder]

Someone already tried that in the past, ended up making their own clone just like noodles bc it didn't work.

[@sutsuj6437]

You could probably do that, although that would require a lot of work. The real problem here is the "nearest obstacle distance" input, which the game doesn't directly provide anywhere in memory. "IsFlying" on the other hand would be simple enough with tools like CheatEngine.

[@mira_nekosi]

@@sutsuj6437 i'd just give the AI almost unchanged picture from the screen, probably with lower resolution and converted into YCbCr

[@airtoumfake]

Dang! Distance to the next obstacle must have been a really good metric! They made it so far so fast!

[@elijahiscozy]

I've been watching you for quite some time now. you deserve way more than 32k subs, your coding is phenomenal!

[@RealNostalqia]

Damn, amazing video! I love watching the AI "evolve"

[@justarandomwoof2247]

One of the rare channels where I have notifications enabled for! Your videos are really cool!

[@theRealMaMo]

Such an underrated youtuber amazing video keep up bro 👍

[@Sans20xyz]

I kept on forgetting this wasn’t the actual geometry dash! Amazing job, keep up the great work!

[@mazenyasser7208]

I'm studying Deep learning and Neural networks and play GD as well so this was pretty fun to watch and learn... My experience is still limited as I'm still a beginner but this inspired me to continue studying it to maybe build Neural nets like this in the future. Keep it up please.

[@azuritii]

man not even the ai can handle that triple spike

[@BNnnen]

you should make a part 3 and give it harder levels

[@BritishBoy]

I second this!!

[@marcespano2866]

Just dont add back on track tho, might be too hard

[@snowleopardseal]

and different gamemodes

[@notphantom_4710]

(yoo I'm in the video) The new version is MUCH better. You are very talented and I love your content. One pet peeve I have for GD clones is that they NEVER get the ship physics/rotation right. But I'm assuming it's very difficult to code. Keep up the good work 👍

[@WojtekXD-bx7jb]

8:03 That leading player's straightfly was insane lol

[@noah_qt]

this is way to advanced, nice work dude

[@stevenoct11]

S-tier username

[@Nsquare_01]

@@stevenoct11 by Nisha

[@spectralkat1597]

im trying to learn how to program and i would like to say, thanks for inspiring me and keep up the good work!!

[@sentricityy]

been waitin for this video to come out

[@grahamb.s.637]

you should try to teach it to play randomly generated levels.

[@CodeNoodles]

That's a creative idea 💡

[@theRealMaMo]

There is just one catch if the level is completely random impossible sections would occur so it would be smart to have prepared sections wich will spawn randomly and gets connected easily so that in a ship part suddenly a cube part spawns and in every part that spans could randomly generate some changes so it doesn’t get repetitive so fast

[@anthonywong7906]

Just a few things I noticed: First, there’s no coins, although to be fair, would be something else you would also have to code, which if it was me, would to be lazy to do. And second, I noticed the ai had some trouble with the 3 spike jumps. If that’s the case, it could also be a good idea to hard code the timings for the more difficult jumps so it doesn’t need to take as many attempts to learn it.

[@CodeNoodles]

Good points!

[@indoterror8886]

It's about learning not following orders

[@SpykoYT]

actually the problem isnt triple spikes its the fact that one it learns the timing for one type of jump it uses that one timing for ALL the jumps so if i made 2 spikes one space and another spike [forcing the ai to jump early] the ai will learn to jump early but because of this will never be able to perform a triple spike

[@dompan9169]

@@CodeNoodles or maybe just stop it from changing the path that already works. There’s no reason for it to mutate during a part it has already completed. It only needs to change just before the part where it previously died.

[@Radu_-ro6fp]

i think i can make some kind of history of their quest to beat the triple spike so 20 cubes randomly started going and trying to beat a level, which seemed impossible for few generations. eventually, after 2 generations, they successfully got to 12% of the level. but, everything changed when a brave cube successfully beat the jump of the 12%, and exploring way furthermore. he got around 3 times further than everyone else, and people were shocked out. eventually, he learnt them about this, and how to do the big part of the level, but there was one problem... THE TRIPLE SPIKE! many considered it impossible, so they retreat trying to jump over a little spike. and much much more generations later, 4 people got to the triple spike, but only 3 made it over it. then a guy just died. and then right at the end, another flew into a wall. but one brave soul, successfully beat the level, making everyone shocked and being a legend to this day.

[@CodeNoodles]

Well put!

[@The_octogamer]

Your avatar makes me hungry. Also great video man

[@unstoppblekid]

remember that straight flying (I don't think you know what it is) could look off because of the ship gravity (just helping). also keep up the great work.

[@Xevf]

Honestly, very cool! You should try(if you want) making it look at actual geometry dash, recognize obstacles and finish level. Would be very cool to see tbh.

[@2to5Raccoons]

Every single time you upload, I look at it and think "Oh, a Code Bullet upload!" then I look closer and realize it's actually a Code Noodles video! I swear one of these days it'll be the other way around

[@alessandroscotton2510]

Super cool!!! Good job man👍👍

[@Eagleleadsmc]

You should make a level editor and release the full thing so that people can experiment with what the AI can beat

[@user-fs8th1pg4y]

I think it will get downed because he stole some assets.

[@Sur3sT]

I dont think robtop would care too much

[@xDLiLi1337]

Please continue working on it, it's amazing!

[@elliothyson9154]

It's cool seeing a person try to make ai play video games.

[@hightierpeasant8048]

he's like code bullet... but edible!!! great videos by the way 😊

[@CodeNoodles]

Hah, thanks!

[@MajesticUC]

This video was very fun to watch. Very cool (:

[@thomasprados9876]

I'm actually pretty surprise that a neural network this simple can perform this good in a few generations, good job ;

[@yannrolland4193]

Very surprised too, only distance to next, not even next height relative to player ... Surprising

[@minibrawl6733]

I waited do long to see part 2 of this

[@woweeeeee]

The polargeist and back on track icons were the most intelligent ones

[@jakinijeninja]

pls do part 3 with other levels i love this!

[@Lord_Nodus]

Cool, I also thought about making a neuron that will play in the GD, but I wanted to make it play in the normal version using screenshots from the screen. But I did not find a normal way to quickly get screenshots, and my programming skills were then insufficient for something other than Python, + the genetic algorithm is too long, since you need to run a level for each check

[@Damiean_]

I hope you get payed for this massive work.

[@TheFinnish1]

It was really suprising that the AI got through (most of) the new challenges first try, I believe if this generation was put into Back on track it would do pretty well!

[@baphimitisgoated4674]

I think the reason it was struggling so much was because the ones that did better at jumping did worse at flying a way to fix this could be by making 2 separately categories of players being both flying and jumping and putting the best of the 2 together as the goat

[@booklagecodwarrior]

Geometry Dash has a great hack that I think you know, Mega Hack v7 which has a lot of variables that you can see and a show trajectory mod which could be usefull for programming ai. It also has a macro recorder/replayer so you could maybe use that to record the best ai runs. Great video I enjoyed it!

[@drippymissouri]

Spikes in gd have a small hitbox so you can go in the spikes a little bit. It makes triple spikes less evil.

[@beariley03]

I watched the first part right on the day you made part 2 lol 😂

[@professorcube5104]

amazing video as usual

[@GigaChadApple]

It's funny and interesting how the AI was struggling at 63% with the triple spikes jump, because I'm sure that a considerable amount of people that played geometry dash cost that part.

[@sutsuj6437]

I think the giving the AI ray casts that can detect obstacles would be more interesting than just the distance to the nearest obstacle. I'm pretty sure these 2 simple inputs are nowhere near enough to play the later levels of the game (Expect if the neural network is so big that it can overfit, but that's lame).

[@helloypiple3281]

I wonna see how AI will try other levels in geometry dash!!! It's really interesting

[@FlexNiko]

interestingly while i was watching that footage at 3:00 i was thinking who codes like that, no changes, line by line and actually thought u might have setup some script or extension to do this for you 😂

[@AndyGun11]

5:38 this AI was absolutely goated despite being only the 3rd generation

[@ThegamingHouseSUB]

Can you teach these things in like private classes and stuff? This is so cool!!

[@filipharvancik7052]

That flying part gave me so much anxiety, I cant even explain the amount

[@StormSB23]

Very Cool!

[@POGnator]

Honey wake up a new video of CodeNoodles dropped

[@noobofnoobs6165]

Loved the video! Btw, I noticed that the click/output indicator in the neural network visualizer didn’t seem to match up with the “gameplay.” For example, after the ship part it looked like the red was indicating when the A.I. *wasn’t* “clicking,” instead of the other way around. (Also, don’t think of the video as a hot mess, but more of a bunch of intertwined noodles :P) (omg I feel so smart for using so many big words)

[@CodeNoodles]

Yeah, as I said, once the prediction is greater then the threshold the AI will jump but sometimes it doesn't blink because the next obstacle is too close for the value to drop enough.

[@howlinghellgar2214]

The squad rollin up to beat Stereo Madness

[@granitestorm8206]

“With the basic idea out of the way, lets code this neural network” Proceeds to write more code than I can even process with my smol brain

[@iver4268]

It’s so interesting when flying that it will fly down to support blocks but it’s flying. Would be cool to add a reward for going faster

[@CoderCatGG]

Yes, my comment is in the vid Also, could you code speed portals, wave and ufo and then add slaughterhouse (Top 1 right now) and let the ai improve on that?

[@primeslash]

hoping that he adds a part 3 where it gets even better, also im sure you guys noticed. the AI is barely making it in especially at the ship part, it didnt die by a pixel. this might seem bad or scary, but i think its actually a great trait to make it extremely more precise than a human and get through parts like that except super tight 6:05 is a good example, but it also just doesnt look right lol

[@loganclarkplays8717]

Please make more parts I wanna see the end of the game

[@Md5offical]

Math Class: The rest of the class: Fails The nerd: 5:36

[@pancake891]

Best sequel ever

[@Heliomation]

Can't wait to see it try demon levels ~^w^~

[@Dergy]

I just wanna point out that one of the icons you grabbed was just the default mini icon

[@ikilledaman]

Really cool

[@DrablyOlive]

6:31 I don't know why but that one going straight up into the ceiling cracked me up

[@strangerguy323]

Really great vid! but you could have improved it that when the Ai of the last generation is the only one alive it resets, because it has no mutation so it will just die at the same place as before

[@stuff7838]

You should make it where whenever it dies, it tracks what it has seen, so then if it tries demons (bloodlust), it can predict the level, and read it at the front pixel

[@soggycat_]

Good video take a dallasgmd seal of approval

[@memetech-]

This is really cool, but I feel like it is missing a few things… not saying it *needs* all this, it is amazing as it, but if you feel like making it a touch better I would like to see these maybe: Ball mode, ufo mode, wave mode, small mode, upside down mode, advanced gameplay, pads, orbs, vertical gameplay, reverse portal, teleport portal, community made levels, speed portals, practice mode, triggers (which are unfair), coins, dual mode and my favourite - bad hitboxes

[@black_m1n825]

So recreate the entirety of Geometry Dash, including 2.2 update.

[@memetech-]

Tried, failed. But again, an AI in GD is super cool, even if it is technically not even a tenth of GD

[@stormRed]

I'm beyond impressed

[@jacabo136]

seeing AI learn something stupidly hard wave levels would be an experience.

[@Henrix1998]

How is it possible that only the horizontal distance to obstacles matters and not vertical?

[@krauq]

he didnt specify how the input data is given, and i assume that its given like "a spike at height 4 is approaching ahead in 4 blocks"

[@liamwhite3522]

In Generation 11, at 8:43 a lot of the ships fell down onto the spikes but didn't die.

[@pravinkhotare4379]

And you used c++ to achieve this, respect!🔥

[@davdaboss]

It would be cool to do this on some other levels like back on track

[@HelloHSR]

The Polargeist icon is goated

[@zaneinsam649]

It’s a nice AI you created there and over all that you rebuild the first level! But I have noticed: You hard coded where obstacles are, right? It’s fine and all but it would have been cooler if the AI would have worked by scanning the surrounding tiles or send some arrays in the direction of movement. I would understand why you have decided for this method becose it would have been way more work to do my ideas and it must have already taken so long to create the geometry dash (it would be to much time for my lazy brain I have to say), but if you had decided for my method you could use the AI also for other levels just by copying the maps and not more, also the AI would be able to decide for their own path and the most interesting thing: the AI would make more mistakes.

[@CodeNoodles]

The maps aren't hard coded. The only thing that is hard coded is the networks architecture that's all. If the AI is good enough I can modify the map while it is playing and it adjusts automatically.

[@argon1324]

I think what Zan may have meant is that you coded what counts as an obstacle for the AI to measure the distance to, as a property separate from the tiles. Where you put this hard coded obstacle is where the AI would die if it didn't jump. This allows the ai to not have to figure out where to jump itself based just on the tiles, and in fact the tiles don't have any baring on the ai's behaviour. So the ai just counts down until the next death point is x distance away, and then jumps.

[@bluemax6668]

The only input used when not in the ship is the distance to the nearest obstacle, but in the code it just simulates the player for 180 simulations steps (while not jumping) and and the distance used is the one bewteen the player position and the furthest safe position of the player, so it just learns to jump x steps before dying and for the ship its the same thing but with an other value of x. So the neural nework and the evolution process is useless in this case because he could have hard coded it directly to jump x steps before death, that also mean that with this neural nework if you add for example two blocks to block the end bottom path at 6:09, it could never learn to finish the level because the simulation would say that there is a safe platform if doesn't jump.

[@leihau]

While your clone is with no doubt really cool, I'd love to see an AI for the real game

[@u1aq]

mad underrated

[@X3m.Gaming]

i gues now that first level is beaten fun part can start part 3 should be intresting or these series will just turn in to remaking the game and that would be funny

[@remiicatboy]

Cool now make it play Back on Track

[@tequinlake1091]

I was wondering if you were using a program to write the code before you said it xD

[@xav8930]

awesome video

[@assolutogamer9418]

Try more levels maybe demons and let the AI run for some hours or days?

[@kingstoncassell2501]

Awsome

[@HebbexProducciones]

very awesomw tbh

[@sqkriwvy]

the way ai plays as the ship gave me anxiety

[@funnyrockangry]

wil you make it play a harder level maybe? electroman or something?

[@katelynrogers804]

Did you program “obstacle” points into the level? I notice that the input for this doesn’t seem to be comprehensively based on a specific screen or environment input but instead on a preprogrammed trigger. This AI then was mostly just evolving to get the timing right, as since it’s inputs are already so sanitized. If you ever do a part 3 it would be cool to make a “sense” system which takes a screen with block cords and types as input and generates the “distance from the next obstacle” signal itself…

[@TacoPC]

Cool