I Coded an AI that Plays Clash Royale
Vložit
- čas přidán 23. 02. 2020
- This was made when Clan Wars first started and you only needed clanmates to farm crowns.
Watch how I setup Clash Royale on my PC with Bluestacks to create a Python Bot that plays Clash Royale for me.
My other video about Eve Online : • How I coded an AI to m...
Follow me and support me, check out these links :
Support me on Patreon ► / ombarus
Join the discussion on Discord ► / discord
Twitter ► @Ombarus1
Instagram ► / ombarus1
Website ► www.ombarus.com
Email ► ombarus.dev@gmail.com
Github ► github.com/Ombarus
Music ► www.bensound.com
My name is Ombarus and I'm a programmer / game developer from Canada who moved to Japan and decided to try my own luck with game dev.
Attribution :
Extract from Serral vs AlphaStar Game 2 : • AlphaStar vs Serral - ...
As for the timer you can make your own and look to match the game timer. I didn't play a game so it's just an idea that can work. :D
One more thing as for the health of the tower you can use the same thing as you already did (check the color of the pixels at the place of the health bar). You ca check every few sec a few pixel and base of the color have a ruff estimate of the health.
I could start a timer as soon as I detect I'm in the action phase but I would probably be off by a few seconds depending on when I updated the screen. If I want more precision I still need to read the clock at least once. Though using my own timer after that would be a good optimization!
@@Ombarus OK then then check the pixel on the timer spot until it register the right one and then start your own timer and stop the checking process.
@@Ombarus can u make a bot for subway surfers
Genius, I'm gonna use this in my clash royale bot
I was expecting something more complex (with reinforcement learning), but I loved the video and how you explained the whole process with all the problems you got and the way you solved them... comprised the reason you haven't done something more complex! 😂
Really a good job!
I really enjoyed your video, the problems you encountered and how you solved them, I would love it if you continue it.
Here are some ideas / solutions I've found that might help you:
-Each troop has its own sound when placed, so you may identify too many of the opponent
-you could understand what kind of troop the opponent has positioned besides the sound also from the speed
-You can keep track of the time with an internal timer that starts at the beginning of the game
-you could get a personal database of which cards are the strongest against others like arrows and minions
Where can I contact you? I have an idea about game
I really enjoy that u are saying mana instead of elixir x)
Says to me he's a gamer of MOBAS like us
You can use the lenght health bar to at least know the class of the card (tank or suport) and for the towers you now when they changed from full to damage, so after that you assumed that if it doesn't have any heath it is because is down. You can mix the health bar technique for trops and the spaw clock to imrpve the chances of knowing the card. Other way to know the tower heath is to calculate how much dameage has to recive with x card for y secons. There is an optimization for finding the cars in your deck once they do a complete cicle you can predict then.
yeah ive been working on one and it can actually defend and decide what to do depending on what the enemy does ive been working on it for a while and ive been having the same problem with not being able to know what enemies are in the screen, i used the same thing with using the little box to see where enemies are and i had a pretty good player make me a deck and he also helped me decide what to do depending on where the enemies are on the screen, for that i split the game area up into 12 places 6 on the enemies side and 6 on mine so i had a general idea of where the enemies are and i then made 5 choices from best to worst so it knew what to do in every situation, my idea of seeing if the towers are broken was to take a picture of what there is when the tower is broken and look on either side for that and when is sees it it knows the towers are gone, i have some questions about this and i would love some help with trying to make mine better, thank you and great video!
awesome, good job!
You can further improve your pattern matching time by resizing you images to lower sizes...maybe like 80*80 pixels or even lower. Maybe you can convert the crops to B/W and then do pattern matching, thereby reducing one channel altogether from the image. It will make you images blur or B/W and loose on details, but I think that the characters in Clash Royale look very different and small blur images or B/W images would give good accuracy nonetheless...
I don’t understand why this guy doesn’t have 1 million subscribers he deserves it
Instantly subscribed! Keep up the great work with Clash Royale, I am thinking to do something with the informations available in Royale API. basically I would like to make a program who learns from the games of the PROs and helps me understand what's the best answer to specific enemy attacks.
i wish deepmind or some other big ai company made a machine learning version of this, seeing as they made one for chess (alphazero) especially seeing as clash royale is essentially real time chess
It would be pretty easy to figure out how much time is left
there's this framework called tesseract - it does a pretty good job at extracting text from images if you parse the image a bit to make it black and white
you could also maybe look at how much health the units have left based on how long the span of red pixels above them are
this was very interesting. Woud love more AI videos
so much work wow !
This was impressive work. I wonder if this wouldn't be easier by using a video camera pointed at the screen, and you make a program to teach and A.I as you play, something like shape discrimination.
You could use OCR to recognize the time text and convert it to a string. Same thing for tower health, which you look at after checking the bar exists by checking if the color for the bar is correct (otherwise, it's full health).
Recognizing the enemies might be harder, but maybe you can take some pixels below the red level indicator & check the resemblance of the pixel colors to recorded values (i.e., the values of pixels for giant's head add up to around 1000). Besides that, maybe you could do some post-processing on the difference of images & set a certain threshold, so only large & enemy-shaped(vertical) patches remain.
Anyways, interesting idea & great video!
love this video going to sub & like :D
Awesome!
This is amazing
for tower health u can look up the level of the enemy tower and hardcode the max health after first hit make ur program read the health written and once the tower is destroyed enemy level disappears of u know tower health is 0
For number recognition, you can use tesseract ocr in python
maybe you do not need to figure out how much time is left, you just need to know, when the game starts, then you start the regular timer on computer?
Wow great video
MaximillianMus Nr. 208 agree oh yeah yeah
Hey nice work, can you tell how much FPS you got and in which interavel do you take screenshots?
It seems like you should be able to train specific AIs to accomplish specific subtasks like elixir detection or unit detection or time detection.
It would be easy to train those to be robust to changes in color and placement by image manipulation of the training sets.
In fact you should be able to train a single network to output each state variable you're interested in with individual and combined losses.
And while I'm thinking about it, those subnetworks could be directly fed into whatever network architecture you're working with.
"easy"
@@Seedzification yeah I take your point. 'Easy' in this field means it's something that has a solution that's already known. It still takes effort to implement, but 'hard' doesn't mean 'takes a lot of work' it means 'it's not clear if you could even do it'. Elixir detection, unit detection, and time detection are 'easy' tasks only in the sense that anyone who does machine learning would likely be able to accomplish it (even if it takes a lot of work to do).
as in for the position of the units I think you can use a bitmap and as in for unit type you can find out the static addresses of the units using cheat engine, same for the timer. Great work tho
advanced level you can fully use a proxy packet sniffer to analyze the data and then make an AI to react with the specific packets but not sure if that would be legal :D
This is awesome! Great work!
Can you use movement speed to be able to somewhat vet which enemies get which counters? Usually tanks move slower.
Clash royale does have an api that you might be able to get some data from, such as grabbing the deck of your opponent so that you only have to look for those cards. You sadly can't use it to battle or anything but it may have some useful info for you to use if you ever revisit this. You might be able to change the pixels to black and white and make reading the screen easier too.
I tried B&W but I found it easier to say spot mana by simply looking at which pixels has the most red in the bottom instead of trying to use the relative luminosity of a B&W picture. Not aware of the API you're talking about. Do you have some documentation for it?
@@Ombarus Yea, developer.clashroyale.com, you have to make an account to see the documentation for some reason but theres a bunch of data that can be grabbed from it
How long would it take to calculate if you give the entire game arena as input?
In the game, the time for the 3 crowns win is the first 3 min and the time for the sudden death is 2 mins just after the 3 crowns time ends. I guess the game time is not a big deal but time taken by the CNNs to recognize the characters and processing is the real heck. Predicting the mana of the opponent team and placing the right troop in the correct position is what pro players do while playing the game. Perhaps, mimicking that strategy could be a better idea.
Yes, It would be nice to know the time precisely but a good hack would just be to keep a separate timer... it might be off by 3-4 seconds but that's probably fine... Predicting the opponent's mana would be relatively easy too if I could identify the troops on the board... so it all comes down to identifying troops on the board.... I'm pretty sure there's some convolution networks that could do it already (after all there's neural net capable of recognizing thousands of images).... I just don't have the knowledge to make one from scratch only for clash royale :).
This video deserve a three crowns
thanks
GG dude GG
Wow you're genius
The easiest way i think is to inject into the game code and directly check the game memory for objects
I heard about this but I have no idea how to pull this off, especially for an Android game running inside an emulator.
For the mana you can calculate it Your self Ande for the cards after you play a card it goes back in the cycle so you can esactly calculate, after you ha e played 4 cards and seen which card you ha e at hte begginning, all the card that you ha e during the entaire game without cheking the screen
Which software you use
I think there's an API now
I’ve NEVER seen someone automating any supercell game…
Great video!
Guys, what if you use sound to identify minions, every unit has its own unique deploying sound or several, u could match them, and the time number of level appears, to stick it to them, so that you could calculate speed per square and where it will be at the certain moment of time?
you can spectate yourself on a different account and see what cards there playing and where and how much elixir they have
Can it make for every game ?
Hey bruh
Divide image of game on cells that would help and also you can lower size of image and AI left only to do
hello,your vid is very interesting😂i want to konw if the chatgpt could
help this work
I may be able to help you with that, you got the source code somewhere?
I do have the source but I haven't used it since the video so it's quite outdated at that point github.com/Ombarus/clashAI
What module did you use for the machine learning
I didn't get very far with Machine Learning. I was trying to implement reinforcement learning with SciPy but it's not really made for that. I tested some simple algorithm using tic-tac-toe and couldn't get anything to work. What I show in the video is a simple hand-crafted algorithm
3:23 lol as a clash royal fan and a magic the gathering lover I can relate :)
You should code with the 2.6 hog. You basically can defend any unit with the gunner and the snow golem. Perfect for attack and defense if you know the Strat and works with the program to make it simpler. Skel+ gunner. ice spirit+ gunner Log + gunner all works
Oh my God just no.
I hate 2.6 hog cycle
This is not an AI,this is just a bot. Clash Royale Bot
It would be so nice to have a clash royale AI like in chess, that would beat the best players with bad matchups
wouldnt work, its a matchup based game
Good work! I am impressed. But I think that even if you will get all the information about the current game board state and your and your opponent cards on the field and in the waiting stack and remaining towers health and time it still will not be enough to write RL algorithms. Since you need to model behavior of each of the cards in the present and the future. So you need to know the exact algorithms of each card behavior implementation in all different possible situation. This info Supercell has but it is very difficult close to impossible deduct from analyzing the screen. You can try to go in the route of DeepMind with Atari games and try to learn algorithms solely from screen pixels but these were much simpler games I think.
It would be more realistic to do such a work, like writing an algorithm for ideal play from inside of Supercell. Possibly they even have something like this, since they have a training arena, right?
I don't think it's very important to know the behaviour of each card because you cannot control them once they are on the board so it becomes much like a game of chess. For any set of time/card/position/health that you have seen before the ML should be able to predict the outcome. Of course, I never got there so it's only a theory
Well usually in chess and other similar games to choose a good move you do look ahead (min max or monte carlo trees), you try to predict a combination of moves yours and your opponent until some depth and at the end of the branch you estimate the received position and propagate back the conclusions. So beside of predicting what card opponent can put on the game field and also where the algorithm needs to predict how existing cards on the board and towers interact in a given time frame. Seems as very interesting research problem in case you have access to all game info.
Can you make for moblie legend?
Can you try again whit the help of ai
I want it for clash of clan as well
you can make a current version, thanks in advance!!!
link for the AI plz :)
hey ombarus can u code an AI that plays clash of clans
bro it would be much better if you directly use data from the source in code form in backend rather than in visual...backend code will clearly tell what is what
If you know of a way to decode or capture data from the game playing on an emulator I'd like to know how to do it because I can't think of a good way of doing that.
I think there is a java decompiler that splits the entire game and its java code
Any one named not a bot or not a ai makes it damn obvious its a ai
Think about how the recording work. Grabbing image and doing analysis is slow. Too much input out put; if you perfectly simulate what will happened in the battle filed then we could just grab the packages for troop placing. Furthermore, knowing battle system Behind is the key to train a strategic AI which could predict win or loss.
Base builder bot
Can u make a bot for Android
I'm creating my own AI for this game with about 80 lines of code, and it got to 253 trophies, but I'm struggling with identifying where the opponents are. How did you do this?
1. did u ever find a fix?
2. could i use the ai?
After *sometime* I made myself (AI)
Now I don't want to do anything ✌️
3:01 so pretty much egolem
Yes -_-''. But I don't think my AI can reach Arena 11 so Giant-Balloon it is!
In reddit they selling clash Royale autootouch bot
From what I see it's a bot to open chest and lose a certain number of trophies... it doesn't actually try to play the game.
@@Ombarus I don't understand why people pay for that 🤔🤣
@@Ombarus I don't need to pay to lose trophies Im already a pro losing trophies
🤣🤣🤣
@@eoalmadanepai4683 The tilt is real... I was trying to make a AI that's better than me.... I failed but it was fun :)
Why you dont use Api of clash royal ?
I'm not aware of Clash Royale having any sort of Api. If there is please let me know where I can find some documentation about it!
Is there any source code
It's not very pretty but you can check out my github : github.com/Ombarus/clashAI
I can just say wow 😀😍
I don't know if you implemented it, but you can use a Nvidia GPU to do the pattern recognition with openCV. This is probably possible with a modern computer but is far from a weekend project
3:23 MTG player? 😅
Code pls!!!
I committed whatever I had here : github.com/Ombarus/clashAI Use at your own risk.
Just leave it on overnight and you will probably be in a really high arena
mana = elixir??? xD
The pink stuff that you're always missing 1 for the perfect defense....
Apparently, he says Mana for people who might not be familiar with the game, yet they still clicked on the video because they are interested in the AI topic.
Sir please contact I need your help
Get Elon fund ai for this game.
He scaired of ai yet he want a chip to be put in our heads.
Can you make an ai play a roblox game called black magic 2
So it's not really AI, just a bunch of ifs. A little disappointed , but good effort nonetheless.
I did spend hours trying to program a Deep Reinforcement ML using sklearn but the library doesn't really support it and it didn't work so I had to leave that part out of the video :).
@@Ombarus I'm trying to do the same, but with a different game, and I've been struggling to implement the ai for days. I might have to go with a simpler approach too.
I liked your video, but calling it an ai is misleading.
@@machomb hey i was preparing myself to create an AI for a game.I am completely new to AI and wanted to do something like shown in this video but for a different game, i can code in python and other languages too could you just give me a head start of what all things i need to study and know so i could work my way up
Thanks
Why not just make a CNN instead of hassling with this?
That was the idea. Never got there though. So many questions. How to get enough training data? what kind of CNN? How do you train it? It's not like Clash Royale provides an API for AI to play like Starcraft and I can't exactly make my own simplified game like Chess or Go...
@@Ombarus yep that is where the challenge lies. I will try clash royale on my own. You would have to fiddle a lot with game data and modding the game perhaps. If you are looking for a fun project we have this discord for rocket league discord.gg/H4zQfUQs most of the bots there are hardcoded but some are machine learning and we are working on a gym env that has all the bugs out. Sort of a beta version is now out. Let me know if u joined. Good luck and have fun with future projects
@@Ombarus The hardest thing is to find a good loss function. How do you know whenever playing a card at a certain position is good or bad?
@@trickybarrel444 I had a few ideas. Of course rewarding damage to the enemy's towers, punishing damage to your own tower and damage to enemy units. For enemies it might be enough just to reward any decrease in the number of enemy units on the board.
Coded and AI should not be in the same sentence lmao. You don't code AI, AI self learns. You coded a bot: not an AI. AI also learns and gets better over time - which your bot does not do.
Make coc bot and give the code!
I don’t know if you would ever be interested in doing this, but you can make ridiculous money by coding AIs for games and selling them. Leaguesharp was just a script service and they made over 10 million in their whole lifespan