AI Learns to play Geometry Dash || FINAL
Vložit
- čas přidán 10. 10. 2022
- You guys wanted it, so you got it! Let's have the AI play some more levels from Geometry Dash!
In my last two videos I have programmed an AI to play a clone of Geometry Dash, and in this video, I have updated it to support some new levels. If you like this kind of content, make sure to subscribe for more.
LINKS
▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀
Support the channel: ko-fi.com/codenoodles
Itch.io: codenoodles.itch.io/
GitHub: github.com/OfficialCodeNoodles
█▀ █ █ █▄▄ █▀ █▀▀ █▀█ █ █▄▄ █▀▀
▄█ █▄█ █▄█ ▄█ █▄▄ █▀▄ █ █▄█ ██▄ - Věda a technologie
The way the AI handles the ship never fails to make me nervous
Same
Imagine a video of just the ship entitled "All your anxiety in one video"
@@KryllyxOfficial Hehe, lol.
Same
It’s like cube but flying
The AI passing Dry out in the 1st gen is amazing.
did better than i have and i have done it over 100 times
@@landon2plantsyoure a human and its a neural network. do you see the difference here?
@@clyde-or-monty yeah but that neural network hadn’t evolved and adapted to the level yet, meaning that A.I., without evolving, still had more skill with only the knowledge learned from completing Poltergeist, Back On Track and Stereo Madness, tho i do get it, those levels provide with everything you need to pass Dry Out other than the basic upside down part
@@clyde-or-monty we also have neural networks in out brain
@@DeamonDragoon yes the reason AI has neural networks too is because it’s supposed to be mimicking intelligence/ humans
Funny how “once distance to next obstacle is close -> jump” is really all you need to do the majority of the level lol
well, most of these beginner levels are just piss-easy timings, no fakes, no complicated parts with orbs etc.
Lol you think this how stuff is made, neural networks require a lot of calculations and mathematical equations and calculus, you are talking about simulating an entire brain with gradient descent and back propagation algorithms, it's not only distance to next obstacle, he might have said that to make things simpler for viewers or else he wouldn't require any machine learning algorithms,it's more complicated ma friend.
@@cinemacat4108 dude you can see the node slowly turn red the closer it gets to an “obstacle” (which I think he just marked all the obstacles as what it needs to jump over lol)
@@cinemacat4108 🤓
@@cinemacat4108 what do you mean by simulating an entire brain? You can see the entire neural network in the top left.
Thanks for your hard work! I have a suggestion: To help improve watch time of your videos, commentary during periods where the AI is “training” may be helpful. I do not make videos but it is something I have seen in the past, so feel free to take it with a grain of salt. Keep up the good work 🤙
Thanks for the suggestion!
@@CodeNoodles I am also guessing you have seen "Code Bullet" and how he does his videos considering you use a similar style and soundtrack to him.
@@theautonomous the "soundtrack" is just non-copyright music, not really super interesting that he uses songs that sound similar
@@Solutra true lol never occured to me.
You should totally add the coins
Watching AI learn their path towards multiple objectives would be cool
Waiting for part 4 when you just recreate the entirety of geometry dash
ai beats vsc
AI beats Bloodlust
AI beats TON 618
@@jwjustjw8946 woah no way an ai can beat the biggest black hole
@@zahrannahyan9666 that was a joke right
My favorite part was when he said “it’s code’in time” and programmed all over the place
ok, this is starting to get stale
@@ThatNoobKing no
top ten most original comments
@@garbageman543 It's'eryn't'ood
Después de muchos anos de chivo comment
I can’t believe he was like “more?” And we said “More!” And he went “MORE!” AND GAVE US SO MUCH MORE STUFF
This really shows an interesting point in neural net development. If you use inputs that are extremely bimodal where it's nearly always in its best interest to act one way or another, the AI very quickly tunes it correctly! I think having the vertical check was very useful because in nearly all situations it is a bad idea to jump when something is above you. It is also nearly always useful to jump when an obstacle is directly in front of the player. It's really interesting to see how the AI nails this game once given the proper tools to succeed! Thank you for sharing this
The reason its hard jump over triple spikes in your game is because the actual game's spike hit box is a small square that's on the tip of the spike
its not on the tip but more of around the middle
And also robtop squishes the spikes in a little, to make it easier
@@undefined06855 only on one triple spike in stereo madness
@@brandonoverbey6948 i know it also happens on cycles
@@brandonoverbey6948 every triple spike in the robtop levels is squished in
crazy how ur able to upload so frequantly with this high quality videos
Thanks so much 😊
@@CodeNoodles remember me when you get famous
Yeah :D
i thought the series ended. good that they did not
edit: it would be awesome if you could release this project as game so we could experiment how the AI would try to straight fly (it does ship parts really weirdly compared to normal GD bots)
8:36 yellow portal is really weird when rotation is not changed
9:00 transition perfectly syncs
edit2: i did not watch whole video before commenting lol
i see you everywhere
Literally saw the 2, video before 10 minutes on my start page… pressed the subscribe button and then… CodeNoodles uploaded a new Video. I very like these Videos pls do more of AI and Neural Network Coding. Thank you
No, thank you!
@@CodeNoodles fu-
AI Warehouse has infrequent but great AI content
Would also be interesting to see how the ai behaves if it doesn't get reset after each level, so it could truly be learning
8:13
This is interesting. The neural net and task are so simple that one of the randomly generated nets of the starting pool was able to just do it first try.
That's pretty dope.
Absolutely my favourite series :DDDD Great job CodeNoodles!!
I loved all 3 videos so much, keep the good work!
I hope your channel blows up. These vids are fun to watch
Your hard work and productivity is amazing!
Cool video! I wasn't expecting it. What would be cool though is if they had coins. Like in the generation after it (if there even is one), they would try and get the coin, because they would be awarded.
I think it would be interesting to see how the AI from one level reacts to a different level. While they are mostly the same mechanics, in theory there might be some differences.
Maybe a round where all the successful AI are brought into a brand new level and then see if they all win still. Ofc not saying you need to do it if you are done with the project, but I think it could be fascinating.
Discovered this channel recently and I love it
I love how much better at Geometry Dash the AI with 8 brain cells is than me.
next you need to take the keystrokes and see if the actual level works. :) love your vids btw
I know a part 4 isn’t needed, but I think watching the AI figure out how to play a memory level would be cool. Something as simple as 3 paths and only 1 can let you pass or something a little more complex like the level marathon. Great video 👍
Edit: I did not watch the whole video before commenting
If you make further parts I think it would also make sense to try making a level as hard as it can possibly go in this clone
CodeNoodles is a secret agent working for noodle companys because every time i watch a video i get insainly hungery and i am currently boiling water to make ramen... thank you
7:18
Damn, this AI plays like everyone.
i was always going up
@@doshi6311 up is for the coin
I've been waiting for someone to do this :)
Please continue this series until the ai beats bloodbath, also I think you should commentate on the ai's progress to make the video more entertaining
I can't believe you just made a new version of Geometry dash! amazing!
The music syncs perfectly with dry out, I am at peace
how an ai can be better than me in dry out :(
nice video man, you need more recognition, your videos are too good!
the sync in back on track is sick!
Nice! Time to watch some anxiety inducing flying!
:D
This AI Beat back on track in 59 attempts, Im scared of what it can do...
ahahahahahahahahahahahahahahahahaahaahahahahahaaaaaaaaaahahahahahahaahaahahahahahahahahahahahahahahahahahahaahaahahahahahahaahahahahahahahahaahhaahahahahaaahaaahahahhahahahhahahahhahahahahahahahahahahahaha
get it? because its impossi-
The AI finding the most dangerous way to complete the ship:
*Walter White cooking GIF*
Same icon beat both back on track and poltergeist, he is an MVP
certainly
Hi CodeNoodles,
Awesome video! I wanted to know what those middle hidden neurons are composed of.
-Thanks!
Those neurons are called hidden neurons. They don't represent any of the inputs to the network but rather act as another layer of computation for more complex behavior. They take in the outputs from the input neurons.
This is awesome 👏
Every video of you makes me hungry like fr that bowl of spaghetti looks SO GOOD 😆😭
Nice ! Now lets see if it can beat Clubstep and Deadlocked
Is it possible to put the winning AI of each level against each other to see which one can adapt the fastest to new levels?
bro yessss my favorite geometry dash CZcamsr bro ur so good
I find videos of AI doing perfect runs very relaxing
Lets go, new upload
This ai beats the levels faster than most real people that play the game
Yeah and it handled upside down gravity like nothing lol
This is really cool, but you should make it do Clubstep next.
Do you think it would be easier to just try and integrate the bot into the actual geometry dash?
probably not since its made specifically for this clone and he doesnt have Gd's source code so he can't exactly do anything about that
Yay New Ai Teaching Geometry dash Video
yay finally a new part!
You'll be famous if you continue with this project man
I feel like he's nearly reached the end of the things he can do in this series
it would be amazing if you could train your AI on your clone and then hook into the real game and play the levels with your pretrained network for part 4 :)
Part 10 : AI Beats Clubstep in Geometry Dash
love ur job btw
I feel like at this rate, it's suitable for part 4
that one guy: decent enough. time for bloodbath!
I'm actually not surprised they beat Dry Out on the first generation. The AI always seemed to either get really good, really fast, or get stuck on one part for a while (like in the first level)
To make making the levels easier just make it so you can copy the levels from the game and drag them into your own version or have an external program change the files to the correct files to transfer the levels just to save you time
try adding in the coin paths for the levels (you don't actually need to add the coins). i think it'll be interesting to see if the ai goes for it or no
I'd love to see it handle a demon level after all that. Maybe without all the fancy graphics, but seeing it overcome a really hard level would be very fun to see.
noy try time machine (lets see if a AI is smarter than me XD) also great vid
Very cool stuff
Quick question: if we record the AI playing a level and want to show it to you, how would we do that? Do you have a Twitter we can tag? Maybe a community post we can link videos in the replies to?
Good question. I will be making a public discord server soon so that would be the perfect spot for any clips.
@@CodeNoodles Ooooo, nice! Can't wait
I thought it's gonna be in the real game but this is still awesome
This is actually so cool
I’ll be honest I want to see this beat a level such as clubstep, but that could lead down a rabbit hole. Who knows it might lead to it beating slaughter house, but that would take a lot of work on your side.
you should reverse the rotation of the cube in the reverse gravity section to make it seem more natural
wow, adding that vertical collision detection really improved things, huh?
Idea: Create a skilled Ai on more complex things (blue orbs, ball gamemode, etc) and try to let it beat insane levels like xstep.
Now add a neural network to make it alto adjust and learn every time it misclicks as well as letting it know when it has less than 10 seconds left on the clock so it heightens the accuracy req to not misclick then.
Oh wow the AI fluked Dry Out - not to mention the couple of seconds of perfect straight fly. I reckon this could be trained to beat demons tbh - with the issue being you'd need to make said demons. Is there really no way to get your AI to work in the real game, because that would be really cool.
that's pretty cool. it's a pity that there are only two game modes, whereas in the original there are 7 of them
AWESOME!
can't wait to put this against acheron
Nice job
Nice music taste
Making the A.I play Bloodbath when?
I wanted to use the level editor but it feels very tedious. You can only grab the tile you are hovering over and cannot cycle through the titles, or at least I can't find a way to. Is there a button for cycling through tiles?
Mouse wheel scroll.
Hello, could you clarify something for me?
So I don't know if I understand backpropagation well enough, the thing I currently doing is
for every layer i--:
for every weight i--:
//output = neuron that the weight contects to
A:= -(target[i]-output)
B:= output*(1-output) //for ReLU
C:= neuron that the weight coming from
weight -= Learningrate*(A*B*C)
bias -= Learningrate*(A*B)
You could probably do this in geometry dash itself. There are some open sourcing projects for it and the modding community is very large, so maybe try doing this in the actual game?
cant wait to see it pass slaughterhouse and limbo!
Please make this record the series of inputs given for the final result of each level so they can be ported into another software that can replay them in GD and also try more difficult levels like Clubstep
yeah but the problem is that the physics isn't exactly the same, so porting it over to GD would have problems
@@aadenboy True but im sure they could be refined enough to work for most levels
In the next part you might want to reverse the cube rotations while inverted.
If you ever make a new part, that is.
I'm waiting till it can beat firework. Also good job I could not learn this in a millon years
Cool! Now make the ai beat bloodbath
You should make AI to learn some extreme demon. That would be amazing!
Yo because of your videos I got into AI programming and just recently made my first neural network with 1 neuron, LOL!
Can you tell me where to start and what programming language?
@@guest50035 So the way I started was by watching AI videos to understand the concept then, I went into Unreal Engine 4 and just used trial and error to make a working system. I know this is not the best way to do it but I hope this helps!
I can't wait to see if the bot can play Slaughterhouse
Damn, man's good at straight flight
id love it if you made a cheat that reads the completed percentage and the objects in the level instead of having to make your own GD, i think its possible with dll injections. it seems fun
A problem I noticed, I play GD a lot, and if you make a part 4, would you be able to reverse the way that you rotate when you are upside down, the jumping and rotating while upside down makes me... eeeeeeeeee
Could you let the already trained AI do demons? Mayve start with Clubstep and go up to Tartarus (Maybe without the deco because that would be too hard to make)
At this rate you'll have them beating Sonic Wave while I'm still stuck on B.
Up next: Can an A.I beat Slaughterhouse?
I hope it will happen
Up next: Can an A.I. beat sonic wave
You are that one guy that would be able to make a fully working AI in the actual game.
With triggers
Is it possible to make an ai that all that it knows is that it can click, hold, release or not click? And just have a general visual input of the game with no more information?
Like an ai that simply sees the level and has as a goal to get as far as possible, but doesn't know anything else, that has to figure out how everything works including jump orbs, pads, the ship, etc.
IDK about coding or programming AI (or coding in general lol) so idk if it would be simpler or harder to make the ai being able to just see the entire screen and nothing else and go just with that information and learning from there.
It sure would take way more time for the ai to learn to play, but I think it would be even more interesting. And yeah I know, that's very impractical but that would be very interesting to watch.
imagine how funny it would be if he made a part 4 and he made the AI play slaughterhouse
do you mind releasing this project so we can experiment with it and watch how it will perform on harder levels?
great video, one thing I want to say is that the ship just dips too fast
The fact a clone of the game is needed for this just shows how amazing RobTop's coding is
no it just shows that you cant easily extract the source code of the game
1:20 void sprinkler
1:22 pee pads
1:26 confetti sprayer
I cannot wait to see a ai beat blood bath