Exploiting the Rocket League AI Better Than Me
Вставка
- Опубліковано 20 чер 2022
- I am only Champion 2 in 1v1, yet I crushed the supposed Grand Champion Bot Nexto. But how did I do that?
You can play against Nexto yourself to try and apply the tricks: rlbot.org/
Creator Code:
ROCKETSCIENCE
I earn commissions when you use the code to buy in the Epic Store, or RL and Fortnite in game shops. #EpicPartner
Support me on Patreon:
/ rocketscience
Follow me on twitter
/ halfwaydeadrl
Join my Discord:
discord.rocketscience.fyi
And you can follow me on twitch :)
/ halfwaydead
----------------------------------------------------------------------------------------------
Rocket Science
#RocketScience
#RocketLeague - Ігри
I like the style of this video, blending Matt Parker, Numberphile, and your own funny spunk to make this absolute banger! Keep it up!
Thanks! Will do!
> and your own funny spunk
Hmmm... that must mean something different in the coumtry where you live.
Funny Spunk.
I think I found me a new alias on steam =D
Homie talking about another man's spunk
@@RocketScience On the offchance that you take suggestions, could you do a dldsdr/dsr explanation/ comparison on this channel? would be great! Works way better than anti-aliasing for me atleast.
Awesome video! Glad to see someone exploiting the weaknesses, and exploring them so in depth. Might have to link this video from RLGym twitch just for the neural network explanation. Training these weaknesses out of a bot is difficult in self play, but the RLGym community is certainly going to try. SSL bot 2022? We can hope! 🎉
This was a very great presentation of the basic concepts behind Deep Learning and Reinforcement learning for a larger audience. Really, you did the impossible and distilled it down in a comprehensive and concise way. Also this video was very well done, great narrative and editing. Your best video yet!
More videos like this. This would easily become my favorite rocket league channel. Very entertaining and informative.
Great video! I was hoping to see something from you on Nexto.
I found that one way to defend Nexto’s nasty flicks is to defend from the back wall above the ball. Since Nexto doesn’t really defend the back wall, it doesn’t really know how to get around a defender playing from that position.
These bots are a lot of fun/frustrating to play against. Nexto 1:1 is 90% loss ratio for me, playing normal. However, 2:2 where I get a Nexto against 2 Nextos is more like 60%+ on the win side, and a whole lot of fun. Add an Adversity bot to each side for added mayhem. Element and Necto (before recent update) are fun 1:1 (Element is like a Plat level, capable with some shots, but not too good). Element is low enough that I can get practice dribbling without always losing the ball to a bot (like Nexto would be).
Amazing video! Entertaining and informative at the same time! (although I'm probably biased since I study computer science lol).
If I had to criticize one thing I'd say that switching between looking at the left and front camera felt a bit too frequent to me.
Also, I really appreciated the transparent Halfwaydead when explaining how nexto worked(plus the tf2 reference(plus all the jokes)).
Hopefully the algorythm likes the video too. Good luck on future projects!
OOOooooOOOO am LOVIN this high tech fashioned editing and cinematography, yes the vid is amazing but this alone is GREAT to see, love it man, keep plishing it , u have something special cooking here!
Holy shit you've really cranked up the quality of these videos. Kudos to you man
Very tru
Your hard work will not go unnoticed, this is a really excellent video (as always) look forward to more!
Keep it up with this production quality :) you always made good stuff, but this is just next level! You'll become better and faster at it as well.
Wow this was a great video. So a evo algo multiplies, mutates, and the fittest reproduce. But in deep learning we let it act with a reward function, stop after the match ends and see what neurones caused the increase in reward, slightly increase the weight of those neurones, and then let it act again.
awesome video ! tecko is gonna be crazy if it figures out wall to air plays
If anyone can find it’s weaknesses it’s Rocket Science!
wayton too
@@potatononny Most of the higher level players eventually figure it out.
Love the ‘let’s not use a sample size of 1’ thing
The editing in this video is great!
Best Nexto video by far!
The SECOND I heard of the Nexto bot, I wanted to see your video on it. As always, awesome work.
good video, informative and fun. love this channel!!!!!!
That ending lol
Great video man!
One can only dream of the day that Rocket League adds bots to the marketplace/workshop for an easy one click install.
This is the quality content that I really enjoy. As a QA engineer this really hits me :D
Great video, liked the editing.
Another exploit I found is that squishy saves work everytime. Nexto won't drop the ball after starting a dribble, and will flick it up at you
So well made bro
you are really the einstin of rocket leauge.
Nexto's flicks are INSANEEEEE.
god damn im in love with this video and editing
god damn im in love with trox
great vid man
Nice to see that GC bots still drive the ball into their own goal sometimes :P
This is crazy. Rocket league and neuroscience. So cool.
Thanks for the video
dayum son this is sick
Is it possible to train a bot with a different reward function when you start and change the reward function later on? Because that would allow you to make nexto way better.
Also great video and editing. UA-cam needs to train their AI better to show me this right away.
You deserve more subscribers
Well done
High quality content as always nothing against you or your content; I thought it was common knowledge what Nexto's weakpoints are and exploiting them kind a just renders the bot useless to the point where we're right back to the reason for people starting to make these bots in the first place.
Two things worth adding(pointing out for people watching/reading) when he says manage boost what he properly should've said was Boost starve nexto. Air dribbles is a weakpoint for sure and you're missing another huge one - demos. Combine boost starving, air dribble/aerial powershots and demos and just don't get scored on by hardcoded kickoffs and there you go!
i have a question brother is it also better to play wired on ps4 not just pc and why? even without the polling rate change because i am on console
I played 3 matches against this bot last night. I am a diamond 1/2 in 1v1. The bot scored ≈16 goals or so on me on average every game while the highest I scored was 3...
That last clip looks GC to me.
Any reason why you chose to play the Endo? According to the hitbox visualization videos, X-devil MK2 seems to match the hybrid hitbox better, unless I'm missing something. :)
Really appreciate all the work and research you put in the videos, this channel is the best in terms of facts and intricacies by far
Nexto -> consistency-> GC ✌️
how do you play the bot yourself? where can i find it
You can click Download RLBot for Windows here: rlbot.org/
(Gonna put it in the description)
The most transparent RL creator :)
CN you do more controller test like on razer controllers or did you try the eswap s for Xbox ?? That would be awesome mate.
very good
Yo rocket science does holding accelerate while doing aerials make a difference compared to when your not
1:28 LMAO
Did you have any thoughts about how you might tweak Nexto's reward system to make it better at long term goals? The whole thing must be in a very delicate balance.
Wayton oofers
While I'm really exited by the result the devs got with Nexto, I'd also love to see the journey getting there.
check out rlgym on twitch.
@@kaiyotech I'm watching that stream daily already. I still think creating a recap to show how Tecko learns, how it learns the wrong thing sometimes, so it struggles with unlearning it,would be very interesting as 15-20 minute video.
Hey, i dont know if youre gonne read this and you probably have heard this a thousand times, but ive been really struggling with input lag(or heavy car what ever you migh call it, i know some will say placebo but i played this game for almost 3000 hours so i know when the game feels different) in the past weeks/months, ive tried everything from overclocking my ps5 controller with hidus(which works for a couple of minutes and then after that it gets worse again) to changing the ports on my pc(msi b450 tomahawk max MB) but no success, would a different cable make a difference for example a data transfer 3.2 usb cable? thanks
I think the real question you should be asking is will there ever be a bot better that pro players?
no doubt in my mind there will be a bot that can defeat pros in 1v1... 3v3 is much further down the line as the future planning becomes much more important
Given the importance of mechanical skill in rocket league, I think bots that beat pros is well within reach. Computers are way way more precise and consistent in their mechanics than humans, as we see with Nexto's ground dribble skill that I'm sure will soon translate to aerial skills and possibly passing too.
I just heard that rocket league maps may not all have the same dimensions based on a test on the speed a ball rolls up the wall. I thought this could be a perfect thing for u to experiment if it’s true.
I'd like to see if my friends and I can beat 2 or 3 nexto's. Is there a tutorial for that.
i found the new version harder than v1710.
Also you didn't mention that nexto is bad at avoiding bumps
The new version has jumped a whole rank, I swear. Updated video to follow? 😁
If you do mustys diagonal speedflip kickoff then you can beat it to the ball easily and get a free goal
Tell us you didn’t watch the video without telling us you didn’t watch the video.
@@ogbushwookie5003 show me the part of the video where he used mustys diagonal speedflip kickoff
@@BowOneFire It is one of the three main points he discussed in how to beat Nexto. Starts at 8:02.
@@ogbushwookie5003 he just says to do a speedflip in a normal kickoff. I was referencing the kickoff that musty invented where the opponent doesn't touch the ball at all and you get it all to yourself.
@@BowOneFire Yeah, the one you’re referring to is just a speedflip kickoff. They’re the same thing. What do you think he meant by a speedflip kickoff? A normal diagonal flip?
wow great content. I am a psychologist and just recently got into machine learning. It's really fascinating! Btw am I going to see you at worlds? :)
Thanks and no, I'm sorry. No plans on coming to NA anytime soon. Not in the budget. Also not gonna be at the major this time but probably at future EU ones.
@@RocketScience Damn! Ok then I will see you at EU LANs since I live in Austria xD There will certainly be a big event in central EU again next season
@@EinGuterSamariter Grüß Gott!
@@RocketScience xD Grüß Gott, servus :D bin nur nach Österreich gezogen, eig bin ich deutscher haha
what decal is that
Oh good, I can do 1 of the 3 things.
PLEASE. Make a video describing cross vs circle vs square deadzone. the information is very low online. Some pros like daniel have circle and I dont get how.
Could Nexto or some other bot be trained on the data from replay files from ballchasing?
Nexto has been training by being placed in random moments from pro and ssl gameplay
To answer the question further, replay files lack a critical data point for training bots. The actions. Replay files don't have the controller input that a player took during the game, so training a bot from those replays is extremely challenging or may not even be possible. Further, while there are millions of replays out there, it still doesn't compare to the amount of game play time we can create from letting a bot play against itself. The final version of Nexto had around 500,000 hours if I recall. Getting quality information with actions at a constant skill level relevant to a training bot and getting so much of it, isn't possible with replays.
That hasn't stopped people trying to use replays to help train bots though, as was said, we use replay game states to spawn the bots in so they find themselves in situations they might not have explored themselves, like getting a flip reset in the air. From those game states they can learn to deal with that situation even if it wouldn't have happened naturally.
Lastly, replays can be used for other things, myself and a few others in the RLGym community are exploring using SSL replays to help bots understand positioning and rotations. Essentially creating a "coach" AI that will become part of the reward function.
Can you do a video on bakkesmod delay
what in the fuck, when did bots get even near this good?
That’s the power of exponential growth… you don’t notice when you’re in the middle of it, but in hindsight, it’s blatantly obvious… I wouldn’t be surprised if the next version goes straight to being SSL
What does the v1710 mean? Is it not a version?
Is this the newest version? I was able to beat the v2 legit, but I'd only be able to win against this one through exploiting his AI
Edit: Autocorrect mistake
necto v2 is grand champion in 2v2 lulw
Just let it learn air dribble and try again...
Love your content. I hope there will be a new video soon.
Question: Idk if you ever made a video about this, but there's a video stating that you are faster with air rolling than without. ua-cam.com/video/AkvUa1CC2YI/v-deo.html
Have you ever tested this?
second
fuck
its not gc lvl tho if u just go in air it cant do shit
But it's flicks are better than pro's. Consistency is key and Nexto is very consistent, even if that means consistently bad at defending aerial plays.
Humans are not robots and can't reliably execute the win kick-off/grab boost/air dribble smoothly strategy. That's why he lost to the bot occasionally despite knowing the theory.
Nexto is a very concomplete gc, great in a few things lacking in others, just like some real one's are.
@@chrisziogaming Well i wouldnt say its gc lvl yet, If IT Start aerialing or defending aerials then its gc lvl now its c2-c3
@@chrisziogaming but still so fucking impressive For a bot