Training AI Bots to Fight (they started dancing)

[@cozmouz]

For those curious, Tensor-Board graphs for the final trained model showed "block frequency" way higher than "punch frequency". For some reason the AI resorts to blocking more than punching. I will be sharing more data and answering questions regarding the simulation over on Twitter: x.com/cozmouzz

[@supreetsahu1964]

I know how to solve it @cosmouz make the punch rewards a quadratic function in punch velocity, currently yours seems linear

[@bluestone-gamingbg3498]

@cozmouz whats the music at 3:54 ??? Can i have the title? Could anyone provide me the title?

[@mattstroker3742]

If they are trying to respond to each other's movements also, from the total begining, even before they learned how to stand, then you might be pushing them way too hard.

[@chad_usa]

@@bluestone-gamingbg3498 Ben Elson - 89

[@ssrvendra]

sir, feed them some basic martial arts data and then challenge them with 5 more simulations. see if they improvise

[@amazingnacho2177]

I would say that you need to keep an energy parameter. The reason we don’t see fighters constantly using these dance movements is because it would drain their energy. Add a limited energy parameter and you see better results.

[@pedi-kun3978]

i see any human doing this dance movement will get tired extremely quickly any machine will also wear before get into the fight

[@amazingnacho2177]

Yeah, so hopefully they would learn to min max their stamina

[@philv2529]

But that martial art already does exist it's called capoeira

[@JohnDoe-jp4em]

​@@philv2529Yeah and nobody in MMA fights like they are doing Capoeira... Besides, Capoeira is relatively slow, sweeping movements, not hectically spazzing out in random patterns. A human could do this for maybe 20 seconds before being exhausted

[@rotorblade9508]

yes but they might get tired at the same time

[@gat0anonimo]

Mfs had hundreds of millions of fights just to end up fighting like my drunk uncle

[@wesleysilveira5692]

at the end of the day, thats amazing, its a energized rock that thinks by itself (the computer), and they're evoluting as our species did. if you let them train billions or trillions of fights, they gonna come with insane techniques i guess.

[@bantakkor8039]

Already looks insane ;)

[@PlatinumStrikes]

Nah this was funny 😂 made me laugh out loud

[@mikeblalock4116]

@@gat0anonimo hey! Are you my nephew? xD I'm better sober damnit

[@cosput]

Have you considered that your drunk uncle is a synth?

[@lastyhopper2792]

That's what happened when there's no energy cost for every movement. The fighters' movements simply become erratic.

[@BRAZILIAN_MIKU]

Not really cause you would still be able to be punched or kicked in the face and since your hands are just spasming everywhere you wouldn't be able to block it 2 boxers without stamina limit would be like 2 mike tysons, only in fighting and consistently shortening the distance as fast as possible

[@MystiKhaos]

they move this way because they dont see the floor, only the movements'sopponent. so they dance with each other only

[@KillaAhmadilla]

I think another big factor is they learned to balance. I don't know if they can feel balance the way people can. Their erratic movements are probably to some extent over-corrections.

[@BRAZILIAN_MIKU]

@@KillaAhmadilla that only expains the overcorrection of the legs, the arms spasms is their failure at learning how to throw a punch

[@KillaAhmadilla]

@where_is_sauce You're not wrong. A punch with a bad stance isn't going to be particularly strong though.

[@fractgate6824]

Yellow: shallow and consistent strikes. White: either misses or leaves a crater in Yellow's ribcage

[@Godly_Chair]

fr

[@McCarthy1776]

I didn't see a single strike in the whole video

[@alexbruns5082]

"ShAlLoW aNd CoNsIsTeNt StRiKeS" also yellow (and white): 🤹‍♀🤹‍♀🕴🕴🏋‍♀🏋‍♀🤸‍♂🤸‍♂🤸‍♀🤸‍♀🤾‍♀🤾‍♀🤾‍♂🤾‍♂🤞🤞👉👉🚴‍♂🚴‍♂🏌‍♂🏌‍♂🚶‍♂🚶‍♂🧎‍♀🧎‍♀

[@adsmarket_1]

@@alexbruns5082 went straight over bros head

[@jedimasterjoe5386]

Like Derrick Lewis

[@jaydenstewart190]

Fights outside the pub 3am in the morning 🥶

[@10aelbeg22]

After drinking 1Million gallons of alcahol

[@RRKS]

@@10aelbeg22 Due to the speed, i'd certainly think there's some coke, meth or whatever "up" drug involved as well. Maybe all of them

[@storiesreadaloud5635]

The spastic synthwave dancing is genius

[@deeOOgh]

reminds me of kzbin.info/www/bejne/maKzdWd3Ysd-hKM 😄

[@roux6715]

@@deeOOghI clicked this expecting you to be a bot comment, I was happily surprised, thank you for sharing drunk russian dancers with us

[@nanja7773]

@@roux6715u thought they were a bot and still clicked?

[@roux6715]

@@nanja7773 well unlike bot comments the message actually makes sense as a reply to the original comment and wasn’t just ‘this is the clip you’ve all been looking for’ or whatever

[@ChineduOpara]

That's literally the only reason I watched the whole video... then WATCHED IT AGAIN 😂😂😂

[@midbell]

i am no longer afraid of AI killing off the human race

[@nubbs4753]

You’re fucked in a dance battle

[@Anthonybrother]

Killing us with their awesome dance moves

[@princesaini4975]

The ai watched rumble the movie

[@dtester]

@@Anthonybrother Yeah, we're going get served!

[@RenzitoARG]

Remember a few months ago when AI couldn't draw a hand? Then Sora happened.

[@sirhoog8321]

Fights with your brother be like:

[@rotorblade9508]

they are worse than toddlers 😂

[@wasabininja4749]

More like telling someone with adhd to stand still

[@antonellisamuele6569]

More like your little brother trying to fight you

[@aykarain]

​@@antonellisamuele6569plot twist: he is the little brother

[@Timfoil0]

@@rotorblade9508nuh uh

[@cris89631139]

It might look dumb at first glance but they are feinting each other faster than a human could almost instantly leaving no openings so they have to constantly change their angle of attack and cancel it/turn it into a feint midway through, maybe you could get better results if you train 1 AI to be proficient at attacking an evasive target that could also block (likely with motion capture), another one to be defensive by evading and blocking, and then combine both. Or maybe I'm just wrong and it will be the same result.

[@mattstroker3742]

No, I read multiple tips in the comments thatcwill for sure aid the models in the end, including yours.

[@sectix]

They just need a reaction time. If they reacted a few hundred milliseconds later to each movement of their opponent, they wouldn't do this.

[@BAGELMENSK]

That first thing you said is definitely not happening.

[@BRAZILIAN_MIKU]

They aren't feinting each other because they weren't trained to feint, there are multiple situations where a normal hook would get them but they don't throw it cause they don't know it. When you are about to throw a punch but the person dodges you don't need to cancel the entire punch, you just adjust to where the person is or you throw another type of punch

[@visartistry]

They are not doing shit, they are figuring out how to punch. We know how to punch because we have innate information about the functioning of our bodies. They are not even aware of their bodies. They just move a certain combinations of joint around and if it satisfies some parameters they tend to repeat It. I think you should just leave them training for months and we will see progress.

[@shubhamroychandra5914]

They are trying to balance using their hands not standing on feet

[@bensoncheung2801]

69 👍

[@RenzitoARG]

True, so do we... That is why we balance our arms as we walk and run.

[@ryanoschwald7551]

Yeah gravity needs to be implemented here so that they're More planted

[@rumplstiltztinkerstein]

@@RenzitoARG Are you saying that you are incapable of standing up without using your hands for balance?

[@RenzitoARG]

@@rumplstiltztinkerstein One thing is standing up, another is to balance. there's a huge mechanical difference. You can observe it in babies. They first learn to stand and then to walk.

[@geraldpujaya6682]

I think the reason it looks jittery is that they have the minimum possible reaction time. Meaning that they would react to the opponent's movement as soon as it happens. And if both avatars act like that, it would explain why the block frequency is higher than the punch frequency.. because both of them are reacting very quickly to any movements the opponent makes. Also if you have this in mind while watching the fight, you'll realise how cool these combat robots are, and that humans probably wouldn't be able to fight them in the ring

[@bradfordprice3781]

It takes the human 160 milliseconds to register an opponents movement. The average punch travels 31.68 feet in that amount of time (starting at 0 mph and ending at 12 mph at impact). Part of fighting is learning to "read" the opponent, typically called "fight IQ," by experienced fighters. That's the part that AI will need in order to conquer mankind.

[@Oren_is_tired]

1 mean left hook and they're done

[@daniiii888]

The AI just has ultra instinct, that’s why it can react so fast smh my head

[@Templarfreak]

you have to keep in mind that they have very limited experience and its incredibly difficult for them to learn from experiences with a real human fighter (or hell, even against each other).

[@alwest4472]

@@Oren_is_tiredwait for them to complete;y dodge then punch you

[@micahkent4093]

I wonder what would happen if you were to add a reward of something like “making the least amount of movement, but causing the most impact to the opponent.”

[@thepotatotaxi2430]

Stiff jab spam i assume.

[@aquafox8732]

The "damn bees attacking my face" dance

[@bladeplays7389]

“Bro fighting demons”

[@remipoujoulat7759]

What happens when you want to be a boxer but you're epileptic :

[@ChineduOpara]

...and you end up at a rave😂😂

[@RealKnarli]

This feels like watching 2 stand users fight but I don't have a stand

[@navada4269]

This is a banger

[@alibeknorbekov5963]

Fr

[@Shmethan]

Lmfaoooo

[@ln9593]

It can dance better than me at any rave

[@InnerEagle]

Wait until they understand how capoeira works

[@djashes]

Rawe*

[@hasangarmarudi2178]

We all been watching 15 min to see a snail brain to learn fighting 😂

[@Verrisin]

pretty sure snails have WAY more neurons XD

[@shanewarren8583]

came for the boxing, stayed for the dancing

[@RRKS]

At the beginning of the video i was thinking that the best part of these AI videos are seeing the guys just struggle and fall flat. I was proven wrong with those amazing moves. He secretly taught them to rave

[@user-wd7pp9hx9k]

5:27 he did a spinning kick that was awesome

[@pixeldoor4792]

"A ragdoll without a brain is merely a corpse; so lets give it a brain" Absolute BAR

[@381delirius]

"capable of balancing with ease" *Agent continues to gyrate every limb uncontrollably*

[@ChineduOpara]

But you gotta admit, that last dance-off was SICK 😂

[@amuronam1345]

1:10 “the most complex object in the whole universe, the human brain.” -Human brain

[@user-fd5qz6nv7g]

Born to drunken fist,forced to box

[@babyoda1973]

Sound track was perfect for the silliness 😂

[@ChineduOpara]

It was amazing!!

[@screamingeagle1018]

0:44 "..but to ragdoll without a brain is merely a corpse" - my new favourite quote

[@voidspawn7637]

10:10 this is literally how school fights are 💀💀

[@mhm6421]

4:00 I love how they look like they are dancing, can't stop laughing 😂😂

[@ChineduOpara]

Me neither 😂😂😂

[@ГотаБГ]

Same 😂😂😂😂

[@DiphonFX]

End result looks like a Florida parking lot fight.

[@sgtcole40]

I mean... it might have been a good fight... if it weren't for those damn bees!

[@sampup6871]

this looks so funny its like newborns are learning to fight and walk at the same time

[@CitroChannel]

You might not like it, but this is what peak combat performance looks like.

[@redstartrooper6024]

3:55 what a non stand user sees when 2 stand users fight:

[@jaygorst8121]

8:26 i know what its like to be the gold guy in that scenario. Ive fought silver guy outside a club and just stood there like 'wtf are you doing bro?' While the guy took himself out 😂😂😂

[@user-ks6yp2iy5r]

Cool music makes those random twitchings soo much epicier

[@AdeshBenipal]

wait.. I totally did not realise that you were such a small channel....the video quality is awesome!

[@RayaneAtd]

I thought like he has 200k

[@Tsukiy0mi_moongod]

That's fascinating. Almost like seeing a child learning to walk, but action filled

[@milanstevic8424]

3:52 The name of the catchy electronic track is Ben Elson - 89 (released by Epidemic Electronic). I've recognized this track because Ben Elson's work is used as music for the video game Beltmatic (Notional Games), but normally I use algorithms for this kind of identification.

[@mayuravirus6134]

Dancing WHILE Boxing? Sounds like a mix of Boxing and Capoiera

[@MamaTrixxieAsmr]

Sounds like it but the end result is more spazoeira

[@skitariiranger4346]

You'd think so, wouldn't you, but actually its 2 puppets having seizures

[@AgentQQ8]

WtF … that’s not some new evolved combat technique. That’s two dudes having seizures.

[@Sergio_752]

Good lord, AI learned crackhead fighting style

[@AIShipped]

Great work! You put great effort in making the ai work as well as the editing of the video, keep it up!

[@Oren_is_tired]

Mf looks like he tryna rap battle with yellow at the end of the balance training

[@MkwbeagleIzrahel]

Its like two masters got drunk yet still can dodge while starting to punch

[@doodeedoodee]

7:20 When you and your buddy take the same amount at a festival but only one of you really feels it.

[@thechicken1477]

So capeoira?

[@billwang7891]

How I imagine Jake Paul and Mike Tyson fight be like💀

[@cofftps67yago94]

I love how at the end of the video it becomes a techno dance battle/moshpit of two/drug addicts having a seizure

[@SageMuy31]

Add an energy constraint/input/ or negative reward that prevents too much movement over the course of a few frames. (That is, why dont you try to make the AI lazier?)

[@RealTacticalMax]

Great work 47, your next mission is to stand up.

[@FLAXMS]

Excellent video and thank you for uploading in 4k60fps!

[@spikeytortle595]

"dancing while boxing" bro thats just capoeira but for hands

[@antmoundsock1122]

3:15 wouldnt you want to use the dot product between the punch velocity and surface normal so head-on punches are more valuable?

[@medisty.]

I hope this video gets the likes and views it deserves!

[@Chokymilk.]

“Drunk people try fighting well progressively getting sober”

[@NO_ONE_ZONE]

4:00 why are they dancing lol 😂😂😂😂😂 It would have been funny to put a song there.

[@sevenclovers7]

General: What kinda combat robots were these?! They just kept dancing! Robot: THREE TWO ONE ELECTRIC

[@quez900_4]

1:49 I’m sorry but u said observations complete wrong “ob-o-way-shins” is what u said 😭😭😂

[@CastleRene]

I caught it the first time, but I'm just now realizing it.

[@bantakkor8039]

Looking at the ragdoll model, it has a lot of "joints" for its hands but only two for the feet, our feet are basically hands as well, so it might help with learning to balance if you made the feet more complex as well. Also, we are able to use our muscles with unequal emphasis stabilizing 0-100% on one leg alone, maybe there could be a mass X for the upper body which needs to be stabilized but can be shared unequaly between the legs, with shifting percentages. Anyhow, they need way more standing and walking before they can fight. :)

[@Juel92]

Really cool. They actually start developing some form of strategy and technique by the end. Also still prettier striking than UFC legend Ben Askren.

[@skitariiranger4346]

So thats a joke right?

[@redstartrooper6024]

At around 11:44 there was a frame where yellow’s fist was 1 cm away from white’s head, but it still counted to white 💀, also bro hit a new dance

[@davesmith7976]

New video hype Always keep up the good work

[@RandomGuy817]

8:36 bro was getting aggressive

[@sirhoog8321]

Underrated. Commenting to boost algorithm

[@andyhoughtaling6319]

11:43 - Two drunk dads squaring up at a baseball game.

[@adedaviasnyah5566]

Loll

[@TraianoLiberatore]

This wasn't invented by a.i. Prince nazeem hamed was a master of this technique and many others had similar style.

[@The_official_Roman_Empire]

I bet he gets to the cloud district very often

[@sorensouthard927]

Crazy how they both started jumping the invisible guy in between them.

[@pariodeusex]

well I don't think this model is quite ready for the next terminator but they do seem to have invented a new kind of artform/sport I would call artistic dance boxing :)

[@mr.absolute8556]

the sheer ammount of feints that these two are able to throw to each other, puts all profesional fighters to shame. Truly some high level stuff

[@FurryNonsense]

They're becoming sentient!

[@Aburaishi]

I think a big part of the problem with the final product here is that during balance training, the bots learned to use their arms as integral parts of their balance; the punch training was awkward afterwards because they were still trying to use their hands to balance too much to really ever extend them meaningfully far from their own bodies. I don't necessarily know what the solution is - balance is just kind of a tricky thing to train- but some way to eliminate or reduce the allowed use of arms during balance training would probably allow them to throw punches more freely in later stages

[@Aburaishi]

Changing the training so that it rewards keeping their torso as close as possible to vertical could also help. And rather than random kick force to the hips, adding random forward momentum to the hands (and maybe torque to the hips, for twisting into the punch) could help them learn to compensate for punches, which could translate more effectively to the punch training stage

[@BiggestDuckster]

7:34 what was he trying to do here

[@angel-0f-heaven]

I love seeing the starting frames of a superman punch get interrupted by a Q+E Sieger.

[@bortinme]

seems like they are reacting to each other too fast. human fight they react to each move but ai fight reacts to each line of code

[@dumblockdubbed2455]

they need some sort of energy system, no human would be able to move this fast without getting tired

[@seprious8549]

Dude,I just couldn't... lmao! The setup was so serious and planned out then to witness that mess..omg had me dyin! 😂☠️

[@Aburaishi]

First training session looked like my first hundred Toribash games back in the day

[@MrBioWhiz]

I love the idea of a general AI looking back at this video and realizing we made it's ancestors fight each other for our amusement lol

[@Xeller_real]

I'd love to see a VR fighting game where it places you against a neural fighter. Sick documentary!

[@Drunken_Hamster]

The agent needs a response latency limit both in the perception and the reaction phase. As others stated, it also needs an energy cost that scales with movement size and force. Small movements like balancing or barely moving the arm to feint should cost almost nothing, large movements should cost a little, and consecutive high-effort movements should cost the most and add up to a diminished energy recovery over time. This energy system would be the single most complex part to program correctly, but if done right, it should result in much more realistic movements once you get the training balance right between energy conservation and pacifism. *EDIT:* Also, their reaction strengths need to be imperfect. Lots of times fighters can overcompensate for feints, which then allows their opponent to capitalize with an actual attack, or sometimes even anticipate actions that never come, which usually ends up looking like erratic behavior to the opponent. But these anticipatory moves can also be used to lure the opponent into attacking, which can then allow a fighter to counter-capitalize on an opening that their fake reaction created.

[@finalbreath15]

I love how your accent makes you say "obserwations" and "revards"

[@redstartrooper6024]

8:07 literally beating himself up

[@BeruleAdierant]

This 3D stickman dancing kinda reminds me of Toribash.

[@redacted-tn3tp]

"Today's mission..." Freakbait sounds different

[@kurero1431]

like what others have said. an "energy" cost per movement and force would be helpful, as it limits the movement necessary to keep balance, and also avoid erratic motions. I would also like to add that where a person is hit could affect their balance, especially the head, as that is where our sense of balance is being controlled. we can also observe and learn from our opponent's movement as well, and adapt accordingly. you could train the AI to learn from different fighting styles, give them limited energy, a sense of balance and movements to control it, and have them react to getting hit, or react to movements that would get them hit, sort of like an eye tracking/ object proximity detection. these are highly advanced and complex to simulate. but considering these factors would make you understand that there's more to fighting than movement alone

[@hanivdoniv]

Why is everybody talking that the result being a dance is because there is no energy cost system? Even if you put two pro fighters in a only one round fight or somehow give them infinite stamina, they will still not fight like this.

[@OfficialPhilyblox]

the first training session is basically 2 people dancing without knowing the purpose of the arena

[@ReverseUTTPpolice]

Ai dancing and falling with absolutely no sounds exept for the background music is hilarious

[@bacon_with_brussels_sprout]

The problem with ai is that they will always do the most efficient move that would guarantee the least punishment and most rewards, this would most often result in stalemates. Ai cannot reproduce what humans have, the ability to risk it all.

[@MegaSudri]

the problem with ai are the humans who set up the punishments and rewards.

[@joshuamccoy5407]

It really is so dope, there wobbling because the ai is firing inputs off of both there outputs at a high rate so there struggling to make a precise input that would get them to the goal while not being touched. I know it’s known but to think it’s capable is just fascinating.

[@nemopescante3792]

Absolutely loved the graphic transition between neural network and brain it is graphically awesome!

[@coyotefire69420]

The first training session just looks like a cool dance party

[@ThisGood1234]

10:20 неиросеть настолько точно смоделировала бой двоих пьяных в стельку бомжей около моего подъезда .

[@dallinhunter7385]

How is this music choice absolute perfection

[@shiny4984]

Developer crates AI to optimize hand-to-hand combat, accidentally re-invents Capueira

[@wanderer0013]

Yujiro hanma would stand no chance... We have witnessed the birth of ultra instinct.

[@KeterMalkuth]

What's really interesting to me is that while hundreds of thousands of repetitions is a lot, it's not actually that much in the scale of learning to use your body. Human children doubtlessly take millions of repetitions of any given movement learning to maneuver themselves during play, athletes at a high level making exponentially more by the time they end their career. These models sometimes seem to struggle with simulating complex organisms in movement, but taken in the context of how much effort is required for a child to learn to use their body, even with built in "code" helping them along, the results are entirely to be expected.

[@kidpog3d101]

The gravity is damped too much it should be set stronger also acceleration of musckes should risr linearly instead of "snap"

[@user-tc6ln6ch5c]

3:53 they're so self aware they're constantly falling

[@nuclearmedicineman6270]

I like the yellow fighter's plucky attitude and the strong rearward lean.