A few years ago you helped ignite my passion of machine learning. Since then I’ve learned to code, have educated myself vastly in machine learning. I’m currently building my own small NN. Thanks for opening my eyes to something amazing
There are competitions like that. Not sure if there are any for sports, but I've definitely heard of some other competitions: - Minecraft village/ city generation AIs - Minecraft resource gathering AIs (get diamonds etc.) - Ticket to ride AIs (board game) - bad piggies/ angry birds AIs (not sure which one) - chess AIs ...this is what I remember without actively looking for them. If you look for competitions like that, I'm sure you'll find some.
To get a feeling of how difficult this is: imagine QWOP with 56 keys instead of 4, and you don't just have to run but to play football, in 3D instead of 2.
To get a feeling of how difficult this is, try to consciously control all of the muscles that are required to walk. Or just appreciate the fact that your brain can do that for you while you think about more important things.
I notice that they still seem to move really unnaturally -- their upper bodies seem very flail-y. I wonder if that would get ironed out if they were given some cost to excess movement (just like humans get tired).
I was just typing something similar before I saw your post. I even used the word "flaily". Adding fatigue to the simulation feels like it might have significant impact.
Along similar lines, if there was some cost to getting hit, like being slow to move for some time, I wonder if there would be an emergent consensus to avoid causing damage to your opponent, like an emergent moral code. Would they even develop a tit-for-tat rule enforcement?
They may be moving more efficiently than real football players. People don't always take the most efficient paths when moving, we have to consider extraneous social variables of how our movement looks. It's possible that it takes more energy to restrict the movement of limbs than to incorporate their inertia into the body's trajectory.
@@maelstrom2313 if that were the case, real soccer players would move like this. No one cares about looking stupid if they win (cf. the Fosbury flop, which looks stupid but is standard, because it lets you jump higher). There's also no reason to think that these should be moving efficiently, because there's no incentive for them to be doing so (they don't get tired).
they should do this with real rules (throw ins and cornershots etc), and with 11v11, run it for weeks on different computers, see if they come up with some kind of great strategy, see what formation and stuff they pick.
Even after all these years, you still amaze me with how understandable you make these papers for people like me. Thanks for really spending time in digesting this information to someone who's not in the field or can't allocate enough time to dive deep.
i am a little bit sceptical, with this, there are probably thousands of hours of footage and they show the best most human like stuff, these are things that just comes from the noise. If the ai would be any good they would shoot the ball directly to the empty net every time, then it would learn its better to put one player to the goal
Imagine training an AI to do a task for 100 years in 10 minutes, then exporting it and importing it into a robot to achieve the task perfectly in the real world What a time to be alive!
Your point holds but i dont think that doing things in simulated world corresponds 1 to 1 with the real world. It would still need lots of training in real world (because the simulated world wouldn't hold all the variables that the real world holds and those small inconsistencies add up)
I have always ponder that if our universe is simulated, maybe it is in a supercomputer that does million of iterations, where every iteration takes milliseconds but for us is eternity
*Imagine a Zombie movie/show where the zombies first start out writhing on the ground, and then quickly they learn to get up and walk, and then run, etc, etc*
bruh thats a really cool idea for a game. The longer you live the smarter the zombies get with ai, honestly it would be scary as shit when they all learn to run and look for you in houses. Something like project zomboid but at some point the zombies learn to coordinate.
This actually could be a really cool esports team. Like imagine if madden had AI agents to play the team. It was actually pretty entertaining to watch.
If this simulation gets more accurate, it could influence how football is played in real life. Imagine when the richest clubs can afford to create such simulations and optimise their players.
@@bobisatelier oh yes they would, considering how expensive LLMs are, some SoTA LLMs cost millions just in GPU hours to produce, thats why you only see people like Google doing stuff like this
I would love to find out with a model like this what the ideal play style would be according to game theory and physical limits of the players, but still having to follow all the rules. For example: Would it actually be better even for the goal keeper to move out and play in the field, vs. staying in the goal? I guess humans didn't figure out optimal gameplay yet for soccer and this could lead to new and crazy strategies.
It was my initial thought as well. Using AI we could try so many new strategies. And it wouldn't be hard to give the different AI players different properties to reflect the properties of the real players of a professional team. On eis a faster runner, another have better stamina, a third one have great control of the ball etc.
The talking sound so robot like. Its like you have all the words with your voice already recorded and then you just type the words which automatically creates the voice
I need this to become a thing, like using real football matches to train the AI and have simulated matches between real teams based on how they play. I would love that.
Just imagine the labour curve going down suddenly after a couple of weeks… nobody knowing why… and AI looks like it would start to discuss and talk to each other on the field rather than playing. THAT would give me goosebumps. ;D What a time to be here.
I'm quite curious why many of the graphs have staggers going down, instead of continuously improving. Any idea what that might be about? For example, in "Passing Range" it goes down inthe 1 week range at 6:35
It's complicated because you're working with many variables. Imagine you're going to your friends house but you don't have a map of the roads, just a compass that points to your friends house. You go down a road that gets you closer, but it's a dead end. You have to go back the way you came a little to find the better road that will get you even closer. If we imagine that hip angle is the x axis of a map, and ankle angle is the y axis, then we might settle into an effective ankle angle. But then hey, if we tweak the hip angle a bit we can get even better, but the ankle angle needs to be adjusted again. It's like this, but with way more variables.
Boston Dynamics vs Real Madrid when? Also I want to see 100 vs 100 players and if this can be transferred onto soldiers like in Totally Accurate Battle Simulator.
Remarkable. Ex footy player here, wondering how long till these analyses show how Messi's additional value was in the distribution of labour - a 'team player'.
I hope you have seen "Lawyer Explains Stable Diffusion Lawsuit (Major Implications!)" by corridor crew uploaded a few hours ago. He even uses your catchphrase at the end! What a time to be alive!
I think one of the major improvements in the last few years we've seen in AI training (besides training size/time) is how we train intermediate goals to speed up training. If we simply took the untrained AI with the joint body, and asked them to learn soccer, it would take forever with minimal improvement. But as we've seen in this video, they first train the AI to enable it to walk, dribble, and kick the soccer ball. Once the AI knows how to perform general ball handling tasks, it can use that understanding to more quickly learn to score goals. This might just be my perception though.
the next logical step would be (and it's probably already done to some extent, in one way or another) to split the AI into a "teacher" (supervisory) and a "pupil" (learning) AI, whereby the former one would be in charge of figuring out and setting the incremental goals for the latter to accomplish
This is the kind of niche thing that I personally really enjoy reading in my own time, but it's so much better when you have a PhD holder narrating for you and showing you all of the nuance that you might have missed.
My heart desires a whole tournament with these little AI's. Crazy ragdoll physics and wild flailing bodyparts included. They are so fun to watch! Can you upload more footage?
One day there's gonna be a better way to 'start' these training projects. It just seems wrong that they should start with such little knowledge. I'm sure something will arise as we work towards the future. By the way this was damn impressive! Just the mere fact that it could side-step and through-ball shows how much of an understanding it really has about the game, a surprisingly deep one!
They're training them on rotoscoped humans playing, so was the feints with the footwork introducing "noise" that was creating the initial shaking/quivering? When they're running they take lots of tiny steps, which makes me think that's a latent trait of the human data set foot-faking distorting the training?
Wow, this video showcases the impressive capabilities of DeepMind's AI technology. The ability for the AI to learn and adapt to the complex rules and strategies of football in a simulated environment is truly mind-blowing. I can't wait to see how this technology will be applied in the real world and the impact it will have on the future of sports and beyond.
They're using their body to shield the ball and then turning around to beat the pressure from opponent... Wow.. That's something one of the best midfielders of all time Xavi did often too... Incredible
even more fascinating than usual! this has some extraordinary implications both for AI-generated video and possibly for humans, if 'division of labour' could be translated back to humans in the form of more collaboration. for AI video it surely means we are quite a bit closer to being able to create a background street or restaurant scene and not have to animate every single crowd artiste (a nightmare). this would be hugely helpful in itself, but if/when combined with the ability to direct avatars at a high level (eg "please go and pour some wine into that flute"), it will bring us closer to being able to make virtual movies, especially for difficult or dangerous scenes, which could then be intercut with real actors.
it would be so sick to see full teams progress and having a goalie learning how to goalie and adding in rules like offsides, out, fouls, penalties. would like to see what formations they would come up with or if they would stick with the all forward all back game plan
Physics agents are so interesting, I wonder if there is any demo of this that I can find so I can run it myself. Would love to train goalkeepers and play 6v6 or even branch out to something like fencing!
Just imagine if this ai was put into a robot and you had to play soccer against it. It'd be horrifying with how they move - imagine if they even looked like people 💀 Still absolutely insane progress with this kind of physics-based ai agents! It's crazy!
yeah. I would like to see the rules built-in, as well. Let the AI learn in the context of the real game with out of bounds, and penalties, corner kicks, etc.
@@ChristopherCricketWallace I mean this maybe overdoing it with the rules, but at least have them run more realistically not like they are spazzing out with limbs flailing all over the place lol. Maybe add some cost to limb movement which factors into their stamina bar and when stamina is low they run really slow which would affect their performance and that way AI would learn to be more "efficient". Just throwing some thoughts out there.
And to think how jaw dropping it was to see open AI dominating a 1v1 against pros in Dota 2 years ago. Unreal how far it’s come and it’s only going to grow exponentially from here
How do you train AI like this with UE or Unity? I dont understand how that works and I coulnd't find any tutorials on it but Im not sure what to look for..
I'd like to see a self-driving car trained in this way: We just need to use input data of some data of cars driving perfectly (at intersections, changing lanes, parking) and we can make a simulation where an AI agent is given the ability to press the gas pedal, steer left, right, use turn signals, (etc.) and also given input data about the car's position in lane, positions of nearby cars and pedestrians. We already have the technology to gather this data from cameras (tesla). Then we would train the AI agents. And if they can control these soccer players that have at least 20 different joint controls that need such fine tuning, we should be able to train an agent to drive a car, even have races, do tricks, or just drive from point A to B without crashing.
So so they make all 4 players (2 on each team) learn from scratch? What happens if they drop an Ai in with a team mate that already knows how to play?? Do they learn faster? What happens if they can learn from a team that is better or worse than them??? Sooo interesting. I’d love to learn how to do more of this stuff. I have a few raspberry pi’s that I’m tinkering with but I bet you need a whole bunch more processing juice to do this type of stuff. I hope one day you can make a video that breaks down how to get into a project from scratch or starting from a GitHub library. What a time to be alive indeed. Just ask SkyNet 😉
"What happens if they drop an Ai in with a team mate that already knows how to play?? Do they learn faster? " in principle they definitely should, because there will be quicker and stronger negative feedback = penalty for incorrect behavior this sort of learning isn't really done by copying or mimicking, but by adjusting the neural network attempting to avoid the penalties
If 11 a side and the full rules of the game were added then this could eventually be used to discover new tactics. Similar to how the chess ai developed its own strategies that humans could never think of.
I studied multiagents theory at university and I can't wait to see this expermient go with more players ! BTW if anyone has sources about theory of agents interactions in tthis context
As far as I I understood what I saw, they weren't really passing. It was more like a combination of 0) Run towards the ball 1) Shoot as close to the goal as possible while avoiding the enemy (including bouncing off of the wall) 2) If teammate has the ball then position towards the middle and the goal (essentially where teammate will shoot), which is indeed impressive It may not sound like there is a difference, but there is, because passing to a teammate so he can shoot at the empty net is not the same as shooting the ball next to the enemy, missing the goal, and hope teammate gets there (who gets there, because he expects the missed shot)
As a sports fan, it would be great to see 11 on 11 at a larger scale, add the rules of the game, and factor in variables such as height, weight, strength, speed, endurance, etc.
A few years ago you helped ignite my passion of machine learning. Since then I’ve learned to code, have educated myself vastly in machine learning. I’m currently building my own small NN. Thanks for opening my eyes to something amazing
That is absolutely amazing, so happy to hear your story! Hope the neural network will go on to do great things!
@@TwoMinutePapers You are big my fellow scholar
The first time I learned about them was SethBling'd Mario NN
@@TwoMinutePapers this must be a great feeling to know you are an inspiration for others
@@ThatGuy-kz3fx neural nutwork
It would be cool if there were simulated sport competitions where teams used their own AI model against each other
They do this already with chess, computer chess competitions, the latest winner is AlphaZero-based Leela, just beat the reigning champ StockFish.
@@raylopez99 Didn't Stockfish win?
Loving the idea.
I would like to see this also in FIFA.
There are competitions like that. Not sure if there are any for sports, but I've definitely heard of some other competitions:
- Minecraft village/ city generation AIs
- Minecraft resource gathering AIs (get diamonds etc.)
- Ticket to ride AIs (board game)
- bad piggies/ angry birds AIs (not sure which one)
- chess AIs
...this is what I remember without actively looking for them. If you look for competitions like that, I'm sure you'll find some.
To get a feeling of how difficult this is: imagine QWOP with 56 keys instead of 4, and you don't just have to run but to play football, in 3D instead of 2.
@eetm but your 🧠 is just a processor playing 3d qwop in the dark.
To get a feeling of how difficult this is, try to consciously control all of the muscles that are required to walk.
Or just appreciate the fact that your brain can do that for you while you think about more important things.
@@dzambi I didn't expect this level of existential crisis on a chill Sunday....
Also the buttons are pressure sensitive. ;)
All things should be related to QWOP
I notice that they still seem to move really unnaturally -- their upper bodies seem very flail-y. I wonder if that would get ironed out if they were given some cost to excess movement (just like humans get tired).
I was just typing something similar before I saw your post. I even used the word "flaily". Adding fatigue to the simulation feels like it might have significant impact.
Along similar lines, if there was some cost to getting hit, like being slow to move for some time, I wonder if there would be an emergent consensus to avoid causing damage to your opponent, like an emergent moral code. Would they even develop a tit-for-tat rule enforcement?
I was also about to say this. It seems to be a commonly overlooked issue when training AIs in movement.
They may be moving more efficiently than real football players. People don't always take the most efficient paths when moving, we have to consider extraneous social variables of how our movement looks. It's possible that it takes more energy to restrict the movement of limbs than to incorporate their inertia into the body's trajectory.
@@maelstrom2313 if that were the case, real soccer players would move like this. No one cares about looking stupid if they win (cf. the Fosbury flop, which looks stupid but is standard, because it lets you jump higher). There's also no reason to think that these should be moving efficiently, because there's no incentive for them to be doing so (they don't get tired).
they should do this with real rules (throw ins and cornershots etc), and with 11v11, run it for weeks on different computers, see if they come up with some kind of great strategy, see what formation and stuff they pick.
And add stamina so they have to learn to economize their energy.
And goalkeepers
And a referee.. who always seems to be one sided too for the real effect.
@@letMeSayThatInIrish yeah... genetic, height, weight, heart illness, traumas :D
@@alihms lmao
Even after all these years, you still amaze me with how understandable you make these papers for people like me. Thanks for really spending time in digesting this information to someone who's not in the field or can't allocate enough time to dive deep.
You are too kind - thank you so much! 🙏
It's crazy how the AI just iterates and comes up with through passes, lobs and cruyff turns.
i am a little bit sceptical, with this, there are probably thousands of hours of footage and they show the best most human like stuff, these are things that just comes from the noise. If the ai would be any good they would shoot the ball directly to the empty net every time, then it would learn its better to put one player to the goal
@@DailyCorvid You dont need a ref with robots. That's the point.
I think the pre-training behaviour pretty much captures the behaviour of actual football players
Exactly :-)
I was gonna say 😂
Ah, the bugs kick in sometimes.
From the expert opinion of someone who's never watched football before
@Vixan knowing football, but not acknowledging flopping from a foul? Who doesn't watch football?: you
Imagine training an AI to do a task for 100 years in 10 minutes, then exporting it and importing it into a robot to achieve the task perfectly in the real world
What a time to be alive!
Think i saw something about them doing that with a ping pong machine, right?
I wonder if they inject noise into the simulation to simulate the imperfections of reality.
Your point holds but i dont think that doing things in simulated world corresponds 1 to 1 with the real world. It would still need lots of training in real world (because the simulated world wouldn't hold all the variables that the real world holds and those small inconsistencies add up)
I have always ponder that if our universe is simulated, maybe it is in a supercomputer that does million of iterations, where every iteration takes milliseconds but for us is eternity
@@josesandv More like every millisecond does millions of iterations
I see some serious meme potential in this
We can make a religion out of this
What is the game the AI is playing? It looks like non Americans trying to invent their own version of football 🏈
@@MarcillaSmith Why do you call it football when you bring the ball using hands? Shouldn't it called Handball?
@@LinggarMaretvaCendani No it should be called soccer cuz it socs.
@@LinggarMaretvaCendani handball is when you can't afford a racquetball racquet.
Honestly the first one where all the players are writhing on the ground looks pretty accurate to me.
Seizure?
@@KangJangkrik The other team hit me.
Yes, that's advanced training on what to do after a foul
neymaring the shit out of the game
If that was truly all you need do to play, I might stand a chance of qualifying.
*Imagine a Zombie movie/show where the zombies first start out writhing on the ground, and then quickly they learn to get up and walk, and then run, etc, etc*
and once the movie credits start rolling, there's Pink Floyd's classic playng in the background
"...Hey, teacher! leave them kids alone..."
bruh thats a really cool idea for a game. The longer you live the smarter the zombies get with ai, honestly it would be scary as shit when they all learn to run and look for you in houses. Something like project zomboid but at some point the zombies learn to coordinate.
You voice are a IA too? o.O
This actually could be a really cool esports team. Like imagine if madden had AI agents to play the team. It was actually pretty entertaining to watch.
I see great potential in deadly sport types played by AI.
Worth checking out altered state machine and their upcoming games (FIFA AI League and AIFA)
Nah
@@Ulexcool Hater
@@michaelatorn8380 any examples of what sports specifically, can't think of any other than extreme diving
8:22 **gets lightly tripped over, falls down and has pain seizures**
The most realisting thing in the entire video.
0:53 me when mom says I can't have another pack of fruit gushers until I eat some broccoli
I want more Ai Football!!! I wanna watch a full match! Looks so fun. Imagine what they might do years from now??
It would be cool to watch the 50 days AI vs the 3 days just to really show the improvement.
I would absolutely watch videos of a bunch of ai players fumble around like this for hours.
Me too and I would like to ask where I can find more videos of these
FIFA AI League is releasing in less than a month. You will get your wish
Just go outside man
2:11 "Holy mother of papers!!" hahahaha love it. Great videos and papers 🤖
If this simulation gets more accurate, it could influence how football is played in real life. Imagine when the richest clubs can afford to create such simulations and optimise their players.
they wouldn't need to be rich probably
@@bobisatelier oh yes they would, considering how expensive LLMs are, some SoTA LLMs cost millions just in GPU hours to produce, thats why you only see people like Google doing stuff like this
Two Minute Papers, thank you so much for providing us such incredible content.
realistic human movement that follows the laws of gravity and moves on its own has always been fascinating to me
The way this dude pauses to extend EVERY diphthong. A-and. So-O. No-O. YE-es.
Wow! The first steps of a Simulation World just started, can’t wait to see virtual peoples doing tasks like our own.
HOLY MOTHER of Papers,... Hahaha
Love u Man.
I would love to find out with a model like this what the ideal play style would be according to game theory and physical limits of the players, but still having to follow all the rules. For example: Would it actually be better even for the goal keeper to move out and play in the field, vs. staying in the goal? I guess humans didn't figure out optimal gameplay yet for soccer and this could lead to new and crazy strategies.
It was my initial thought as well. Using AI we could try so many new strategies. And it wouldn't be hard to give the different AI players different properties to reflect the properties of the real players of a professional team. On eis a faster runner, another have better stamina, a third one have great control of the ball etc.
7:32 AI really pulled a Blue Lock there💀
2:20 We used to do the same trick when we were kids, passing over the wall when we played soccer.
0:56: "How much beer do you need"
Him: "Yes"
The talking sound so robot like. Its like you have all the words with your voice already recorded and then you just type the words which automatically creates the voice
I need this to become a thing, like using real football matches to train the AI and have simulated matches between real teams based on how they play. I would love that.
Just imagine the labour curve going down suddenly after a couple of weeks… nobody knowing why… and AI looks like it would start to discuss and talk to each other on the field rather than playing. THAT would give me goosebumps. ;D
What a time to be here.
How epic is it gonna be when a fellow scholar makes a soccer game like FIFA but using these AI's
FIFA AI League game is about to be released on the App Store
I'm quite curious why many of the graphs have staggers going down, instead of continuously improving. Any idea what that might be about? For example, in "Passing Range" it goes down inthe 1 week range at 6:35
It's complicated because you're working with many variables. Imagine you're going to your friends house but you don't have a map of the roads, just a compass that points to your friends house. You go down a road that gets you closer, but it's a dead end. You have to go back the way you came a little to find the better road that will get you even closer.
If we imagine that hip angle is the x axis of a map, and ankle angle is the y axis, then we might settle into an effective ankle angle. But then hey, if we tweak the hip angle a bit we can get even better, but the ankle angle needs to be adjusted again. It's like this, but with way more variables.
@@coder0xff thank you so much for this explanation!
Boston Dynamics vs Real Madrid when?
Also I want to see 100 vs 100 players and if this can be transferred onto soldiers like in Totally Accurate Battle Simulator.
Thank you especially for comments on those graphs. Big respect!
Remarkable. Ex footy player here, wondering how long till these analyses show how Messi's additional value was in the distribution of labour - a 'team player'.
I hope you have seen "Lawyer Explains Stable Diffusion Lawsuit (Major Implications!)" by corridor crew uploaded a few hours ago. He even uses your catchphrase at the end! What a time to be alive!
i love your content bruh, long time fan! ölelés!
The through ball at 4:29 is amazing 🔥🔥
I can't wait to see the Olympic AI Opening ceremony!
I think one of the major improvements in the last few years we've seen in AI training (besides training size/time) is how we train intermediate goals to speed up training.
If we simply took the untrained AI with the joint body, and asked them to learn soccer, it would take forever with minimal improvement.
But as we've seen in this video, they first train the AI to enable it to walk, dribble, and kick the soccer ball.
Once the AI knows how to perform general ball handling tasks, it can use that understanding to more quickly learn to score goals.
This might just be my perception though.
the next logical step would be (and it's probably already done to some extent, in one way or another) to split the AI into a "teacher" (supervisory) and a "pupil" (learning) AI, whereby the former one would be in charge of figuring out and setting the incremental goals for the latter to accomplish
This is the kind of niche thing that I personally really enjoy reading in my own time, but it's so much better when you have a PhD holder narrating for you and showing you all of the nuance that you might have missed.
I like your storytelling, sir. What a time to be alive!
I, for one, welcome our new AI football overlords.
on multi-agent ai, do we have to relearn a strategy for 11 v 11 instead of the current 2 v 2? i assume so because in theory its a different game?
these little foot ball player are so fun to watch i could watch that for the whole day
My heart desires a whole tournament with these little AI's. Crazy ragdoll physics and wild flailing bodyparts included. They are so fun to watch! Can you upload more footage?
One day there's gonna be a better way to 'start' these training projects. It just seems wrong that they should start with such little knowledge. I'm sure something will arise as we work towards the future.
By the way this was damn impressive! Just the mere fact that it could side-step and through-ball shows how much of an understanding it really has about the game, a surprisingly deep one!
@@DailyCorvid loving this
They're training them on rotoscoped humans playing, so was the feints with the footwork introducing "noise" that was creating the initial shaking/quivering? When they're running they take lots of tiny steps, which makes me think that's a latent trait of the human data set foot-faking distorting the training?
0:57 looks like a real football game to me...
4:22 Has a promising career ahead.
I wonder if it's possible to train certain skills first, like getting up, sprinting and kicking separately, then continue with the game training.
In the first seconds: "... this Increadible AI..." meanwhile the AI spazzing on the floor xD
Wow, this video showcases the impressive capabilities of DeepMind's AI technology. The ability for the AI to learn and adapt to the complex rules and strategies of football in a simulated environment is truly mind-blowing. I can't wait to see how this technology will be applied in the real world and the impact it will have on the future of sports and beyond.
They're using their body to shield the ball and then turning around to beat the pressure from opponent... Wow.. That's something one of the best midfielders of all time Xavi did often too... Incredible
The fact that it learned to embellish an injury, only to pop right back up is great
Bro got nutmeged💀 0:27
even more fascinating than usual! this has some extraordinary implications both for AI-generated video and possibly for humans, if 'division of labour' could be translated back to humans in the form of more collaboration. for AI video it surely means we are quite a bit closer to being able to create a background street or restaurant scene and not have to animate every single crowd artiste (a nightmare). this would be hugely helpful in itself, but if/when combined with the ability to direct avatars at a high level (eg "please go and pour some wine into that flute"), it will bring us closer to being able to make virtual movies, especially for difficult or dangerous scenes, which could then be intercut with real actors.
I was thinking teams could learn new approaches or possible plays.
"please" 😂
1:13 This is proof that not knowing how to play football can cause extreme cases of seizures.
it would be so sick to see full teams progress and having a goalie learning how to goalie and adding in rules like offsides, out, fouls, penalties. would like to see what formations they would come up with or if they would stick with the all forward all back game plan
I'm fascinated by the game AI Roguelite and stuff like this video, so I'm subbing now - Look forward to seeing more of your content!
this channel is so underrated. Also its quite scary
Physics agents are so interesting, I wonder if there is any demo of this that I can find so I can run it myself. Would love to train goalkeepers and play 6v6 or even branch out to something like fencing!
I wish there was like a place I could go and watch these AI do their thing in real-time, I could watch that stuff for hours.
Just imagine if this ai was put into a robot and you had to play soccer against it. It'd be horrifying with how they move - imagine if they even looked like people 💀
Still absolutely insane progress with this kind of physics-based ai agents! It's crazy!
you do remember the part where he said that the ai trained without a refree? Gosh you need an ambulance
RoboCup Humanoid League
@@asrar4907 Ahha yeah you'd definitely need at least an ambulance after playing against these
I think they did that already
AI scientists should have a competition to see whose model can train their players better and have an actual virtual football cup.
That red AI trying to feint injury crack me up
it would be interesting if they added some stamina function, so they run more realistically and not swings their limbs all over the place.
yeah. I would like to see the rules built-in, as well. Let the AI learn in the context of the real game with out of bounds, and penalties, corner kicks, etc.
@@ChristopherCricketWallace I mean this maybe overdoing it with the rules, but at least have them run more realistically not like they are spazzing out with limbs flailing all over the place lol. Maybe add some cost to limb movement which factors into their stamina bar and when stamina is low they run really slow which would affect their performance and that way AI would learn to be more "efficient".
Just throwing some thoughts out there.
One of my favorite 'agent' training examples. So wild.
Ok, now put the code into 22 boston dynamics robots and let's go
5:15 It seems that it was trained with Neymar footage haha
It would be awesome if they can also simulate some form of fatigue, some movements look too energetic or wasteful.
And to think how jaw dropping it was to see open AI dominating a 1v1 against pros in Dota 2 years ago. Unreal how far it’s come and it’s only going to grow exponentially from here
1:00 the footballers like babies he crying
How do you train AI like this with UE or Unity? I dont understand how that works and I coulnd't find any tutorials on it but Im not sure what to look for..
*DeepMind - Magical Skills, Goals & Assists - 2023 | HD*
Happy noises from Switzerland!
6:40 its interesting how that curve looks similar to the Dunning Kruger Effect
This guy was an opera singer in his previous life
"the corner of the internet where we look at blue curves and make happy noises" is the best quote ever
0:53 ok thats scarily accurate to how a seizure looks like
The red teammate at the top looks like something out of a nightmare
5:55
"5 years pass by in 3 days. Are you thinking what I'm thinking?"
Well I'm thinking about that Black Mirror episode with a clenched stomach...
I'd like to see a self-driving car trained in this way:
We just need to use input data of some data of cars driving perfectly (at intersections, changing lanes, parking) and we can make a simulation where an AI agent is given the ability to press the gas pedal, steer left, right, use turn signals, (etc.) and also given input data about the car's position in lane, positions of nearby cars and pedestrians. We already have the technology to gather this data from cameras (tesla).
Then we would train the AI agents. And if they can control these soccer players that have at least 20 different joint controls that need such fine tuning, we should be able to train an agent to drive a car, even have races, do tricks, or just drive from point A to B without crashing.
So so they make all 4 players (2 on each team) learn from scratch?
What happens if they drop an Ai in with a team mate that already knows how to play?? Do they learn faster? What happens if they can learn from a team that is better or worse than them???
Sooo interesting. I’d love to learn how to do more of this stuff. I have a few raspberry pi’s that I’m tinkering with but I bet you need a whole bunch more processing juice to do this type of stuff.
I hope one day you can make a video that breaks down how to get into a project from scratch or starting from a GitHub library.
What a time to be alive indeed.
Just ask SkyNet 😉
"What happens if they drop an Ai in with a team mate that already knows how to play?? Do they learn faster? "
in principle they definitely should, because there will be quicker and stronger negative feedback = penalty for incorrect behavior
this sort of learning isn't really done by copying or mimicking, but by adjusting the neural network attempting to avoid the penalties
Ayo very cool video! But why does the "Division of Labour" Chart start so high? They didn't Teamplay in the beginning, did they?
If 11 a side and the full rules of the game were added then this could eventually be used to discover new tactics. Similar to how the chess ai developed its own strategies that humans could never think of.
4:22 that's pretty much Messi's signature move.
I studied multiagents theory at university and I can't wait to see this expermient go with more players !
BTW if anyone has sources about theory of agents interactions in tthis context
Absolutely fascinating. Thank you
If someone makes TMP out of context, “one second passes in exactly one second” would definitely be in it
"AaAnd. NoOow. BuUut."
4:22 The blue guy having a quick panic attack off to the side before returning to the game
As far as I I understood what I saw, they weren't really passing.
It was more like a combination of
0) Run towards the ball
1) Shoot as close to the goal as possible while avoiding the enemy (including bouncing off of the wall)
2) If teammate has the ball then position towards the middle and the goal (essentially where teammate will shoot), which is indeed impressive
It may not sound like there is a difference, but there is, because passing to a teammate so he can shoot at the empty net is not the same as shooting the ball next to the enemy, missing the goal, and hope teammate gets there (who gets there, because he expects the missed shot)
Reinforcement learning 2022 trend: get a lot of data and through it into GPUs training with visualization -> obviously get something look eye-catching
Will any football club like barcelona use it to develop new strategies?
As a sports fan, it would be great to see 11 on 11 at a larger scale, add the rules of the game, and factor in variables such as height, weight, strength, speed, endurance, etc.
At 4:20 you can see they have learned to fake injury, truly mastering the professional level of football.
Before training is like watching a fly breakdancing after a blast of bug spray 🤣