Rocket League Bots Are Now Better Than Pros...

Rocket League Bots Are Now Better Than Pros...
Sign up to The Next Rank - thenextrank.com/
My Rocket League Settings: • THE BEST ROCKET LEAGUE...
Follow my Second Channel: @ApparentlyJack - Rocket League Pro Analysis
Socials:
► Follow me on Twitter: / apparentlyjxck
► Follow me on Twitch: / apparentlyjack_rl
► Follow my Instagram: / jackbenton.23
► Join my Discord: / discord
Music from EpidemicSound and Artlist.io.
#RocketLeague #ApparentlyJack

[@sidabundai]

1v1 them Jack

[@Shabso101]

Definitely, i really wanna see Jack go against Seer

[@the-method81]

I'll take appjack to win.

[@itscmg331]

would be cool ngl, but these bots are EU servers only iirc

[@Petals-kk1lb]

@@the-method81the worst version of seer is around 1400 in 1v1. Could be an upset

[@Petals-kk1lb]

@@Jimmy19382no it isnt💀 Ultrawide GC, the creator Seer showed that it can beat SSL’s in a 1v1. Watch RocketSledge’s video

[@SixFt12]

Seer V7 really shines most in 2 v 2 programming so I'd love to see you and a teammate 2 v 2 the bots! Please make this happen!

[@benjamin8558]

id like to see what having a bot would be like as a teammate. how would the passing look like

[@oibara2]

@@benjamin8558 it looks this good because it's passing to itself, i think - like an intelligence using two cars like you would use two hands. i don't think it'd work half as well with a human.

[@fowlnuke]

The creator said in Rocket Sledge's video, that Seer V7 was trained exclusively in 2v2 as far as i remember.

[@soumalb10]

On Johnny chanel

[@benjamin8558]

@@oibara2 it could most definitely work in the sense that it knows where everyone is and how much boost everyone has. you could probably pass it to the bot itself as long as it doesnt try and steal possession from you. or it'd read your 50s, flicks, backboards, etc, almost instantly

[@jacklewis5017]

25:10 - "they swept Oxygen, these guys can sweep bots" throwing shade at Oski and Joyo 😂

[@2CamBad]

lol yes the comment i was looking for

Grab Leth for a 2v2 against these bots. His engineer brain will help you analise their patterns and he's the best at never forgetting these are code and not humans so he will not waste plays with tactics that dont work against "machines"

[@gameing_0]

100% agree was just thinking this

[@tingledinkle]

Leth did go for a 2v2 against this bot around 6 months ago. That was V6 and the results were similar to Vatira and Atow match against them. 3-2 with the win on Leth’s side. Though they have improved ALOT, I doubt even Leth can win against them now.

[@rinplayz1]

Jack, This version of seer is currently only a 2 v 2 version of seer and won't work in 1 v 1's now there is an old version of seer, seerv4, but its not as close to pro level (closer to SSL level) so not as good. So to play this version of Seer currently, you will need a tm8 to play with, and be sure to both be in a car with at least an octane hitbox or the bot breaks. Hope this helps Jack!

[@MetroTO.]

We really need a seerv7 skill lvl on a 1v1 bot

[@spazneria]

@ 6:42 the bot didn't jump because Atow couldn't touch the ball (he could still touch it even if it wasn't a good touch), it jumped because they learn from playing against different instances of itself, and it has never learned to go for a ball heading to its own backboard in the way that Atow did. They have learned that they can shoot when the ball bounces like that though. So it would have gone up regardless, if Atow hit a psycho there they would have just been scored on.

[@AfonsodelCB]

basically, it visualizes every car in the pitch as copies of himself, and it thought it'd get a favourable position by challenging against its atow self

[@snurffff]

​@@AfonsodelCB It doesn't "visualize". It has just never been in a situation where the opponent has gone for a psycho in that position.

[@VoidLantadd]

​​@@snurffff Bro it's just a word, chill. And he's right anyway. At that instance it is making the most optimal move based on if a copy of itself was in Atow's position. Nitpicking the letter when the spirit is correct doesn't help anyone.

[@snurffff]

@@VoidLantadd I took it as them saying it literally. I didn't think they were using it as a figure of speech. Because it's pretty plausible that someone might think the model is actually doing anything akin to visualization.

[@Phippe]

24:18 They do in fact know the time, and although they’re not required to keep up the ball at zero seconds, they still try to do it every time and regardless of the current score

[@tubax926]

more sportsmanship than most players haha

[@LungsStealer]

they try noticeably harder when they’re down

[@Edss-nr6xw]

That's an interesting insight into their motives. It seems their only concern is scoring and the ball hitting the ground stops them from doing that. If it cared about winning it'd stop when they're one goal up and could possibly go into overtime.

[@Zbro12]

6:18 looks exactly the same goal as when Zen went for a phsyco today in the eu open quals swiss stage round 3 and he barely missed it with Dralii finishing the goal off.

[@neil_surya]

looks exactly like the last 453987435 times ive tried to psycho in comp

[@gnashr4366]

Yeah it's cool how pros are still learning to disrespect pyscho attempts and immediately counter, and the bot just does it the first opportunity it gets

[@Zbro12]

@@neil_surya So true

[@abee127]

For the dota 2 bots, they gave them a 200ms reaction time handicap. I think adding such a handicap would make these bots more instructive.

[@TheyCallMeJinger]

There is currently a version like this being made called ripple. Its training is streamed 24/7 at RLGym on twitch

[@JPClips1009]

1v1 them, also would love if you get another pro and 2v2, maybe on comms

[@hawleyjames7094]

yes this is what we want, maybe retals?

[@Prometheus770]

@@hawleyjames7094 I think rétalas would be best because they’re weak to demos and they have great chem. Other pros/content creators like leth would also be fun

[@Hypnos828]

Karmine Corp has done this already, go watch it

[@Dark.Syndicate]

@@Hypnos828 ?? did u not watch the video?

[@Hypnos828]

@@Dark.Syndicate obviously nottt

[@Pudn710]

I would love to see you 1v1 the bot

[@ImNotBonkrs]

Seer can’t 1v1 jack doesn’t know this yet tho but a 2v2 would be sick

[@sidabundai]

@@ImNotBonkrsseer v4 can 1v1, not seer v7

[@ImNotBonkrs]

@@sidabundai this video is not about seer v4 neither is the rocket sledge video you were talking about you absolute goober

[@slimpickle47]

@@ImNotBonkrs dude, chill

[@Anonymous-mi5le]

⁠@@ImNotBonkrsmate, why tf are you crying

[@kurtmoon6456]

The last touch that the bot makes in game 5 towards the end of the video at 24:25 - Vatira successfully fakes out the bot. You can see that he triggered the bot to jump and hit the ball because he came within proxy. He successfully told the bot "Hey I'm too close to recover" but was already turning around in time to make the save. This is super interesting to me for some reason. Given that they did try to fake the bots out multiple times during those two matches. Great vid, Jack. Excited to see you battle against one of these.

[@yotambronsky9990]

Suddenly I see why they swerve so much. Because it's a way to arrive later at a certain point without losing momentum by braking and then having to accelerate again. Mindblowing!

[@hhaste]

I wish these were available to the public, they would be such great practice in private matches with friends.

[@keatonasuren]

I know, it’s sad cheater ruined it for everyone

[@shavencarrot]

You'd get pissed on

[@madeshsan]

Rocket league can simply add them in exhibition matches, just like the bots that have been in the game for ten years now. Maybe buy them from their creator, but always keep training and updating them 🤷🏻‍♀️

[@oibara2]

@@madeshsan aren't these bots designed to keep learning and improving? i thought maybe it's more difficult than we think to stop them learning and have them stay at a certain skill level.

[@RL_Happy]

​@oibara2 thats something you can easily remove

[@ptuller99]

The bots also don’t wave dash and speed flip very little, something to keep in mind as it’s either not as important as represented in pro play or it shows the further improvement potential for the bots

[@Kyuubiboy321]

I think that when you have the option of having literally no reaction time needed and near perfect positioning, the actual speed of the car to get to where you need to be is less important - because sometimes increasing and maintaining the speed of the car also means that touches, positioning and rotations become less precise, I think

[@MadNoisy]

lol. You can see Jack's confusions. Sometimes I try to think like Flakes and other content creators but looking at these bots gameplay.... pffft crazy man..

[@aliddlldablldoya]

One of the most interesting parts watching these bots on Johnny's channel is that they actually use powerslide a lot for slight sideways strafing while still going forward. Those tiny corrections to put them in the perfect position just scare me.

[@abduxcooper4089]

Amazing video Jack, really enjoyable. In terms of it reaching rank 1 due its consistent high level play, it will, if the developer fixes the issues with it kind of breaking in the middle of matches sometimes. Also the only way the bot learns is playing against itself tens of thousands of times, and you can take advantage of this by just bumping it so many times, it wont know what to do, because it hasn't experienced this in its learning matches. also , very strange to see the bot with ball cam off, cause it would see the ball all the time with the ball cam on ,so you would think the developer would turn it on during its matches with each other .

[@SweetShotGaming]

It always gets a reset because it’s coded to recieve a reward for getting it, nothing to do with a strategic advantage.

[@MadNoisy]

In training they get rewarding but not when the bot is released. So when people get to play these bots or any bot it's not in training mode anymore. 🤘🤘

[@SweetShotGaming]

@@MadNoisy It trains and “learns” to do resets unnecessarily because it learned that it will get a reward if it does. In testing/production, it performs the maneuvers that it believes will give it the most rewards, and hence will continue to do unnecessary flip resets. 👌🏼🤙🏼

[@KCM25NJL]

Regarding the pre-jump for Atow's miss at around 7 minutes, the bot didn't know Atow was going to miss, it also didn't pre-jump "in case he missed", but it essentially looked at the "state" of the environment and there was some number (a probability).... that said something like "Atow has a "17% chance of touching the ball after rebound"..... so it used that "83% chance of miss" as a decision making property for it's next action. We humans kinda do the same thing, but with about 200ms extra lag and no real numbers. Very much simplified of course :)

[@geollert2750]

Great Video and good analysis. 💯

[@Jakeuh]

i like the idea of doing a 1v1 you have a lot of interesting takeaways. I feel like its a double edged sword, these bots might be great for speed/reaction time training. Also i feel like you should get chronic or firstkiller and do the 2v2. There's a whole other element when it's a duo of bots

[@groos3449]

They not only can see time but they start to play different depending on the result. Previous versions of the bot actually were actually programmed to troll if they realized there wasn't enough time for the opponents to win

[@gavinderulo12]

How do you define "trolling" to an agent?

[@groos3449]

@@gavinderulo12 He thought it to own goal and start spinning, etc. Then you give it an "incentive" to do it and a "punishment" if it does it in a bad time, like when they are losing. However I believe this was scraped because they aren't doing it anymore

[@dr_redbanRL]

I also was impressed with that backboard read and the reaction time to score when Atow tried the psycho

[@AfonsodelCB]

I think the aerial plays that confused the bots the most was when the challenging player went backwards or sideways, since it's unlikely the bot tried to score like that when training against itself

[@zaidalyafey]

A few remarks 1. These bots are aware of time. You will notice they became more agresdive towards the end of the match if they are losing. 2. These bots are 2v2 only. So they will not perform goid in 1v1. 3. Regarding demos, it will be a matter of time till they are able to incoporate them. 4. These bots are not programmed deterministically, in simple terms via thousands of hours of self-training, they know what is good vs bad.

[@jameslightforce6194]

I'm guessing it's DRL or RL based then? Do you know what policy algorithm it uses?

[@zaidalyafey]

DRL for sure. Not sure what specific algorithm, though.

[@gavinderulo12]

​@@jameslightforce6194would be surprised if it's not Proximal Policy Optimization since that's sort of the current state of the art.

[@gavinderulo12]

Depending on how the actions are sampled it could still be deterministic, so being in the same state yields the same action.

[@Egrodo1]

Love the bot content, one of the most interesting developments RL has had imo. Keep covering it Jack :D

[@wilfordbrimley6938]

13:35 absolutely. I'm not doubting it one bit

[@neotreo3670]

Something to keep in mind with the bots, doing a flip reset dosent happen naturally it has to be rewarded, so if the bot is doing resets in a seemingly random position, it may just be doing it because it'll get more reward and is in a position where it wont lose the ball

[@danielwilson7494]

Great vid Jack. Super interesting to watch how these bots play.

[@noahd8673]

I believe these bots aren’t programmed to calculate the trajectories of players/balls but are trained with learning methods. So as a result it could learned to approximate the math behind the trajectories over time as it learns.

[@Quiet_3bro]

a pro + a bot vs another pro + a bot in a 2v2 game seems interesting

[@alphasushi9178]

I feel like theres a way to always make sure a bots air dribble bump works, if they get it mechanically perfect, there will be situations where the best possible defender will not be able to save it. What im worried about is if peak rocket league is just whoever gets to the ball first and sets up the perfect air dribble bump

[@kylokat]

2:50 14:55 Base on my understanding on these bots they were encouraged to do as much flip resets as possible during training with a reward mechanism, kinda like how you treat a dog snacks when they do something good. It is likely that these bots have picked up a bad behavior during training that they can be rewarded by doing a unnecessary flip reset without actually using it to its advantage. This is probably also why they rarely actually attempt to use the flip reset to shoot for a goal even if they could.

[@drewgangi]

Definitely was thinking about you the whole time I was watching the video from Johnny yesterday. Please play them. Would love to see you and retals playing them!

[@marongogt]

Regarding why the bot so often goes for a flip reset: An older version of Seer couldn't deal with flip resets, so the author probably gave a high reward (in reinforcement learning terms) if it pulled of one (or defended against one), regardless whether it used the flip. I guess the frequent resetting is just an artefact of that. To be fair this is just an educated guess.

[@theyseeyou7283]

Imagine your own private bot in training that learns your plays and habits then replays the matches it had with you to fix its mistakes and beat you. That would be the one of the best thing's tbh

[@Royalrights]

We're gonna have a Deep Blue moment soon 100% if they keep improving at the rate they have.

[@davidbreton927]

Seer v7 is a 2v2 bot not a 1v1.

[@ImNotBonkrs]

This is what I’ve been saying these commenters ain’t tryna listen to me tho it literally came from the creators mouth in the Johnny boo interview after oxygen 2v2ed them

[@davidbreton927]

@TheoneUnkiejay exactly he also mentioned it in the Rocket Sledge video. I believe he also said Seer v6 trained for 1v1 but im not sure

[@bblvrable]

I would have to imagine that they can achieve the same level of play every time. There's no reason they shouldn't. If they're being fine-tuned through reinforcement learning, they may also continue improving, learning from their mistakes (eg - it would eventually pick up on the telltales of someone intending to demo them, and figure out the best way to dodge the demo). I believe the reason they reset any time they can is because of their reinforcement algorithm rewards getting resets. If the reward is high enough, it will get resets whenever possible, even if it has zero intention of ever using it. It makes sense to set that goal because it does give an inherent advantage when you have a reset.

[@ItIzCam]

Their weaknesses are demos, like you mentioned, but also double touches. I watched Johnnys video and they never stopped a double touch off high back board, seemed to be noticed by the pros playing seer as well

[@neverbeenoutside4963]

You have some really good insights and conjectures about how the bots work. No wonder you're such a smart rl player

[@cheskul730]

its really interesting that this is a machine using cars as tools rather than it being 2 separate players. its way beyond what we know about the game and the human concept of playing it. which is also probably why they dont go for double taps even when it has the opportunity to do so, as it sees the other ''tool'' available. would definetly be interesting to see it develop some more consciousness and personality

[@quinithy449]

Do it man, it would be so interesting to see you play them

[@mattmca7867]

Reminds me of the nike ad where the players beat the alien robots 😂

[@2_Old]

1v1 them, and win. Simple as that.

[@PizzaTheHutt01234]

Watching the full series, I definitely think that if future versions of the bot can figure out demos, then they will be unbeatable. The only other real weakness is flip resets, but I don't think that a human relying on flip reset only offense would be consistent enough to win.

[@7dayspking]

Where's the weakness to flip resets actually seen in these matches? Seemed to have more trouble with double taps really close to the backboard and that Ceiling double tap play.

[@wyliecoyote8065]

2 observations: 1. You would be a great person to have a chat with the bot developer. You can help him to create the truly unbeatable bot. 2. The Bots are barely in supersonic throughout these games. I suppose maybe its not necessary, but isn't it scary to realize that they could be so much more aggressive and speedy?

[@benbryceholden1]

The other advantage these bots have is knowledge of opponent boost levels (assuming that’s true). That’s an inherent difference between humans and these bots.

[@obscuritygamer2556]

omg

[@Azhultin]

It’s easy to determine boost amounts from opponents

[@gavinderulo12]

​@@Azhultinit knows the exact amount. As well as the states and timers of all the boost pads. That's why it's boost management is so good.

[@oibara2]

they're reacting to server information about the ball and player position rather than visual cues, so players are never going to be able to compete with that reaction speed

[@Glagoroth]

Two things. I think Psyonix should have a sort event where it allows this bot duo to queues Duos a limited time a week just as an experiment to see how far it can go. And I think, even though their whole gameplay isn't replicable, it's not completely unenlightening. I have a feeling we will be able to borrows some tactics from them as we continue to watch them more.

[@VoidLantadd]

My bet is that the reason it get resets and doesn't use them is because in training it is being highly rewarded just for getting a reset as the creator's way of trying to brute force it to learn advanced aerial mechanics.

[@groos3449]

When they played against OXG they were playing very different, they scored 3 straight goals air dribble bumping for example

[@online_spy_007]

In Rocket League settings there’s a Spector ball cam setting you can check, I think it’s camera settings whilst in replays.

[@mcrafter_hi4265]

They don't do a flip reset to get that option to use it. They just reset because they get rewarded for every reset they get. So they have been told getting a reset is a good thing, no matter if it is used later or not

[@Brandagbullet]

DEFINITELY WANT THE 1V1!!

[@UnrealChaosHD]

In Dota 2 there was a Before and an After the Bots came into the game... They brought new strats, new ways to control the map, new ways to move, new ways to distribute the gold, etc. They made the game evolve... the players evolved and became better understanding and applying the bot strats one at a time. The same will happen to rocket league.... you see the way they go for low 50 and fight the walls? that will be the standard for pro players within a year

[@mateojamesmusic]

Definitely wanna see you 1v1 them!

[@gavinderulo12]

They aren't made for 1v1.

[@elzen8897]

Of course we want you on a 1v1 against these bots !

[@UltrawideGC]

Do you think you can beat them? 😉

[@sanikku5]

Please set Jack up for a match against Seer >:)

[@theyseeyou7283]

Imagine your own private bot in training that learns your plays and habits then replays the matches it had with you to fix its mistakes and beat you. Is it possible?

[@jnevercast]

the reset is likely an artifact of the reward function for training, the bot is rewarded during training for resetting on the ball during an aerial, likely to try get emergent behaviour of flip resets.

[@alien_zedzz8008]

Having a good day jack❤

[@Alex-hk3dj]

@25:14 "these guys can sweep bots" Freudian slip? XD

[@williamj.c.3574]

if I remember correctly previous version of seer was also doing that kind of double commit team play we see on early game 5... with relative high success rate.

[@ricobico1872]

9:40 "They always does"🤣🤣🤣🤣🤣

[@Turbo-em7pn]

Yeah do it Jack, 1v1 them or a 2v2 with rentals would with comma would be awesome to see 👍🏼👊🏼

[@snekback.]

After seeing the bots' match against oxygen I think they improved even more after this, they didn't get demo'd as much and went for airdribble/dribble bumps constantly 😱😱

[@Darkrat0s]

My guess as to why they reset and then do nothing is that in their training they were rewarded for getting the reset in the air. Hence, they pretty much always go for it.

[@thekraken4265]

rocket league should add these bots to private matches I would love to try to beat them

[@jeffery2799]

If the bot plays with ball cam off, then I would assume that the bot is getting player data location from some sort of behind the scenes data? If so it's not as impressive. It's like if a bot in COD pre-fire headshots a player coming around the corner because the bot has behind the scenes data of all player locations.

[@led3683]

dark would out pace them surely, good video

[@a.g.3029]

We need to see the 1v1 !!

[@rex_otex8746]

the scariest thing is, they already beat pro players and they have still sooo much room for improvement

[@russellhan3534]

just like chess bots, theyre able to play themselves a massive number of times without restrictions limited to humans -- fatigue, "bad" days, or simply human error. They will undoubtedly become better than human players 1 day, and pros could start scrimming against them to improve, since these bots' ground plays and aerials will become so optimized to the point where there are little to no wasted movements at all.

[@sportstj11]

This is awesome wish we could have somthing like this in training be great to get better and love to see u 1v1 the bot. its a perfect timing jump like a Zero reaction thats just crazy. I wana see a App jack Bot lol haha

[@DaveyEaster]

The bots work by creating an algorithm which predicts where the ball is gonna move next, and the best move to make kinda like a GPT model or whatever

[@benjamin8558]

popping the ball up for a pre jumping teammate is gonna be meta. i kinda wanna see how good they'd get in 1s

[@notheories9012]

I think atow can still touch the ball when the car jumps. I think it just hasn't seen the move before, so it sees it as no/low risk to jump

[@theanomaly3238]

7:20 The bot learns by playing against itself and it jumped regardless reading the bounce off the backboard because it has never encountered a psycho before.

[@ghydeon]

go to settings, camera, spectator ball cam. now you can see Seers perspective with ball cam.

[@oibara2]

at first i thought the demo issue was that they only play the ball, not the opponent, but they do. i think instead it might be that they can judge the opponent's capabilities ('they can/can't reach this ball before me'), but they can't judge human intention ('this guy wants to demo me, not play the ball').

[@jiraiya9987]

I've wanted to watch you try playing against Seer or Nexto since a very long time! please do Jack 🙏

[@ripmeme4345]

4:18 The creator of this bot commented, this bot version only works on 2v2.

[@Felix-os5hf]

Yeah the bot doesn't need to air roll aerial like most pros do, it just points the car straight to where it needs to go that's how it's so much faster.

[@ldw9861]

To me, the flip resets look like a way to slow the balls momentum slightly which allows for a pass or drop down more quickly to its bot teammate.

[@frankiealves8221]

Those bots have to know exactly how much boost you have too by now.because they can focus on everything at once, it’ll be able to calculate your boost total.

[@BerkG87]

Would love for you to chat with the creator about all the questions you had during this video. Are bots aware of game time? etc

[@chiarenza451]

I love when my tm8s call me a bot in rocket League I always say thank you for the compliment

[@JGRworldwide]

Love you Jack keep being yourself love the videos always good luck in playoffs!! 🫶🏽

[@justkant]

You could try the director cam to see a bit better the bot pov but you don't control the player you look at IIRC

[@SynsityGW]

i think the flip reset technique isn't actually for the flip, i think they are just allowing the ball to land onto the weak spot of their car to push them down to the ground faster

[@nickc4553]

I absolutely think these bots will be able to become unbeatable. If you compare it to drag racing, a human takes around 0.2 seconds to react to the light turning green. If these bots continue to sharpen their skills, those skills and the free 0.2 reaction time advantage to ariels and most other plays will make it nearly impossible to beat. Great vid Jack, long time viewer 💯💪

[@WindReader]

This is the deep blue moment in RL. The bots are now as good as the best in the world and although humans can still beat them, they will soon surpass humans. Same thing happened with deep blue versus Kasparov.

[@decisionparalysis7856]

I played against a player in the same cosmetics as the bots yesterday, not sure if it was a real player or not now.

[@JPClips1009]

Jack at another point in the series it did get an areal flip reset on net, actually using it

[@zackpayton7421]

ApparentlyRetals 1v1 these guys? or 2v2, that could be so cool as well. This is insane, and there's no way they dont become unstoppable in this game man. that's crazy

[@ZeRuaL.]

is there anyway i can download this bot to practice against in the rl bot gui?

[@derekfranz7660]

What's really crazy is the bots mechanic's arn't even that good. This basically proves that you don't really need mechanics to get SSL, you just need to be faster then your opponent.

[@Knetterkoekje]

We knew this would happen at some point, where the bots would surpass human ability even in such a complex and mostly mechanical game as Rocket League. It still blows my mind how quickly they have progressed though. And like the rest of the comments. Would love to see you play Seer Jack! Also grats on the massive well deserved 200k+ subs!

[@derekgraham8044]

There was a goal in this series where the bot jumped for a save and missed. The bot jumping doesn't guarantee it will win the hit. I think the bots are getting a little too much credit. They are good, but they have flaws. I've been a developer for 18 years and I am working on a bot using the same bot framework. Bots can make mistakes. You pointed out a mistake in this video where the bots bumped into each other and it let to a goal. One thing is for sure, bots always behave with confidence because their mind is already made up and there are no regrets. Additionally, the RLBot libraries give opponent boost amounts to the bots and show opponents through the ball on 50/50s, which is why they are good at 50/50s. They might use knowing an opponent boost amount to know when to jump in addition to distance. I think knowing opponent boost is technically a cheat and so is seeing though the ball (seeing through the ball is probably why Nexto is good and ground flicks). You should 1v1 the bot, but you should ask for practice rounds before filming so you can learn their exploits.

[@vaibhavchavan4691]

Bots don't have memory so I think it is fine for bot to know the boost amount of opponent. All the pros are also good at knowing the approximate boost amount of opponent.

[@derekgraham8044]

@@vaibhavchavan4691 Thanks for your reply. However, I disagree. Estimates can be wrong sometimes and cause poor decisions. In competitive, you cannot see opponent boost amounts, which tells me that the game devs want to force you to estimate opponent boost amounts. A bot dev could use a neural network to estimate boost opponent amount, consequently causing a bad decision sometimes to make the game even. Alternatively, the game devs could show boost circles in the bot's tags so that humans can see bot boost amounts.

[@Julian-zr8df]

I completely agree with this, the main reason the bots are able to play at a similar level as these players is because they have so much more information than a player would normally have.

[@Azhultin]

@@derekgraham8044it’s easy to estimate opponent boost amounts though

[@Azhultin]

@@Julian-zr8dfno it’s because of their reaction time and game sense, they realize they’re beat they turn for better positioning.

[@tommymartin2699]

Someone may have already described it, and I'm not a programming expert I just work in engineering that interacts with programming like this sometimes, so please take everything I say with a grain of salt and do your own research if you are interested. These kinds of bots are not programmed in the classical sense. There is no "if the ball is X distance away jump" or "if I can get to the ball faster than the opponent I go". There are millions of different combinations of all the different variables it has to watch for and as any player would understand there are rarely clear black and white rules that can always be followed. The entire game is improvised and situational, so classical programming just cannot tackle that kind of problem. It's all machine learning, aka learning by trial and error over millions of iterations and attempts. There are parameters it is programmed to monitor (things like it's own position, orientation, speed, distance to the ball, speed of the ball, boost levels, time on the clock, ect) and then outputs it is programmed to utilize (pretty much just the game controls). This set of inputs and outputs (and the connections between them) is called a neural net. It's a very simplified representation of how animal brains work (just a big web of neurons that monitor inputs like our senses and control outputs like our muscles and organs). Machine learning is a process of training this neural net to better accomplish it's goal. This usually involves creating a "generation" of bots based on random variations in the connections of the inputs and outputs. Let's say we produce 500 bots that are all just random connections of inputs and outputs, and call this generation 1. Throw them all in 1v1s or 2v2s or whatever mode you are training for, and let them duke it out. It would look like an absolute mess, they would all just be randomly jumping and boosting and steering aimlessly with no idea what they are doing. But just by random luck some of them are bound to be slightly better than others. After recording the success of each of the bots in the generation, you eliminate the worst performing bots and replace them with new bots that are based on the best performing ones. To ensure you keep making progress, you randomly modify a few connections instead of just copy pasting the best bots. This is now generation 2, and overall the bots should be slightly better. Do the same thing; play them against each other, eliminate the worst and replace with the best, and make some new randomized changes to test new things. Repeat over and over for thousands and thousands of generations, and slowly you build up bots that are able to connect the inputs it's monitoring with the outputs it's controlling in ways that lead to success (aka winning). The crazy thing about this kind of programming is that even the programmer couldn't tell you why they make the individual decisions they make. All the programmer does is set up the neural net (the inputs it watches and outputs it controls) and the parameters of the machine learning algorithm (how many individuals per generation, how much randomness is introduced each generation, how many are eliminated each generation, ect). The way the neural net changes and the way that impacts the decisions of the bot are entirely automatic and behind the scenes. It's a brain at the end of the day, like a literal (virtual) brain. Simplified, yes. Sentient, no probably not. But it's just as abstract and difficult to quantify as animal behavior is. There are clear blind spots like demos because it probably wasn't defined as something it should be trying to be successful at. It knows that it can happen, but it doesn't really care because it wasn't told to care. There are plays that it won't know how to counteract because it's never seen anyone try it before. If it's being trained only by itself, then it won't know how to defend anything that it cannot do itself. But the things that it's insanely good at (fast ariels, perfect 50/50s, perfectly reading bounces or touches) are all things that it can do and are all things that are decided by parameters it is programmed to monitor (car and ball positions and speeds), so because it is a computer it can react 10x faster than us and micromanage each of those parameters perfectly, while we just have to make our best guess and go for it. These bots will 100% be at a level that is unattainable to humans in the next year or two. They are already there in some ways. What is exciting is that having bots like this that are poking around at the very limits of what the game's physics allow will almost certainly unlock mechanics and game strategies that we would probably never discover ourselves. Hopefully it's net effect will be to raise the skill ceiling and instead of having to wait for the next mechanical god to come around and step up the level we can all already see what the peak level is.