Two Minute Papers - 2021-08-29
❤️ Check out Perceptilabs and sign up for a free demo here: https://www.perceptilabs.com/papers ❤️ Watch these videos in early access on our Patreon page or join us here on YouTube: - https://www.patreon.com/TwoMinutePapers - https://www.youtube.com/channel/UCbfYPyITQ-7l4upoX8nvctg/join 📝 The paper "Control Strategies for Physically Simulated Characters Performing Two-player Competitive Sports" is available here: https://research.fb.com/publications/control-strategies-for-physically-simulated-characters-performing-two-player-competitive-sports/ 🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible: Aleksandr Mashrabov, Alex Haro, Andrew Melnychuk, Angelos Evripiotis, Benji Rabhan, Bryan Learn, Christian Ahlin, Eric Haddad, Eric Martel, Gordon Child, Ivo Galic, Jace O'Brien, Javier Bustamante, John Le, Jonas, Kenneth Davis, Klaus Busse, Lorin Atzberger, Lukas Biewald, Matthew Allen Fisher, Mark Oates, Michael Albrecht, Nikhil Velpanur, Owen Campbell-Moore, Owen Skarpness, Ramsey Elbasheer, Steef, Taras Bobrovytsky, Thomas Krcmar, Timothy Sum Hon Mun, Torsten Reil, Tybie Fitzhugh, Ueli Gallizzi. If you wish to appear here or pick up other perks, click here: https://www.patreon.com/TwoMinutePapers Thumbnail background design: Felícia Zsolnai-Fehér - http://felicia.hu 00:00 Intro - You shall not pass! 00:49 Does nothing - still wins! 01:30 Boxing - but not so well 02:13 Learning is happening 02:39 After 250 million training steps 03:10 Drunkards no more! 03:29 Serious knockout power! 04:00 It works for fencing too 04:20 First Law of Papers 04:43 An important lesson Károly Zsolnai-Fehér's links: Instagram: https://www.instagram.com/twominutepapers/ Twitter: https://twitter.com/twominutepapers Web: https://cg.tuwien.ac.at/~zsolnai/
we should try to make a simulation where movement costs them energy to see if they would avoid too many small and fast movements
um no because that's not how it works. stop putting limitations on AI.
@A Reaction cry about it m8 😎
amazing reflex , you know AI depends on Data you provide a lot.
Nah thats what makes them super
But AI is robot. So they have unlimited power :D
The funny thing is that due to their bodies having the same measures, they've learned that cross countering was the best strategy. It would be interesting to see the same experiment, but with different measured characters.
Just one gigantic dude and one stick figure lol
Just one gigantic dude and one stick figure lol
this can become a tool for coaches to come up with game plans
so if one A.I was taller and had longer reach the other A.I. would be encouraged to slip and counter punches, as opposed to just trading punches.
@Zach F. Or set the “goal” to be knocking the other AI over rather than to touch the other one as much as possible.
"After 130 million steps of training, it can not even hold it together"
My life
Lmao
😔
rip
:'(
yeah me too man , after 1 billion steps of training i still suck at managing my life.
Once the AI get so advanced you should save copies of individual behavioural patterns, name them, and start an arena. Maybe live stream fights? Would this not be awesome?
Yeah men that's incredible idea
I’d love to see a boxing simulation where one character has a shorter build or shorter wingspan and see how it adapts to its disadvantage
HEAD MOVEMENT BABY BOB AND WEAVE
@Boyar Valishin LET EM HAVE IT MAC BABY
I'd love to see boxing with wings
Yeah that'd be really cool
Mike Tyson
An important thing to remember with these learning algorithms is that they're going from "less skilled than an infant" to "basic boxing" in *a week*. It sounds like a long time because we're used to computers operating in milliseconds, but imagine going from not realizing you have limbs to walking around and throwing punches in 7 days. That's a huge amount of learning, even in this simplified system.
@Badger computers in general are not impressive or that this specific project wasn’t that impressive?
@the game ranch Computing technology has accomplished things I wouldn't have even been able to dream humans were capable of accomplishing, but this one was a bit underwhelming but perhaps I'm not fully cognitive of the parameters and what's really being overcome.
This is literally the plot of chappie
@DontfallasleeZZZZ why whould you say for 4 hours a day or 6 when u can just run it 24/7
@Eric Martinez or just jump or duck, they haven't really learned that
I think the reason the blue AI kept losing when the red AI fell over is due to over training for one possible outcome. The blue AI expected an attack and only knew how to win when it was being attacked but didn't know how to proceed when not attacked. It overcompensates for the expected attack and then falls over.
Basically it hasn’t been studying the game itself, but another AI’s analysis.
@Comrad pingu
I'm guessing you're joking, but the serious answer is no. The A.I. just does what is going to win it the match in the best (probably fastest) possible way. Though its reactions all depend on what values it gets for inputs.
It looked to me like the red AI was placing its "head" as close to the ground as possible as quickly as possible, and the blue AI would follow that movement with its own "head" which unbalanced it. Why was it doing that? I don't know.
Were the AI's programmed to only be able to perceive what the other was doing via the "head" like a real human, so the blue would tilt its "head" forward to attend to the red?
It would have interesting to see how a purple AI faired against red and blue opponents.
@Soulsphere001 some people also end up becoming a-holes by doing what they think is the best. Emergent behavior I guess
The AI get into an interesting jab war but I wonder if it's because both characters have exactly the same dimensions or are they marginally different? In real boxing, opponents can be different shapes and sizes with variable arm length, height, reach, punch range, speed, stamina etc... If you had two such AI of different dimensions, perhaps the fight would turn out differently as the AI which is getting out-jabbed might change it's tactics completely?
It's interesting as the style the simulation parameters create is quite accurate for a certain type of fighting. I assume the instructed goal is to touch against the opposition's head and not get hit by the way they are moving is very remniscent of points based boxing jabs but more accurately the probing stages of bare knucklefights where the consequences of getting caught by grazing hits is much worse so lunging and swinging back is a solid strategy when you can win with stiff jabs.
Lately I've been wondering about how long it's going to take before characters in video games are able to move realistically like this, rather than using premade animation cycles and other "shortcuts". This video talks about the millions of steps in the learning process, but once the learning has gotten to a sufficient place, can that movement "model" be applied to characters in a game? Is it versatile enough for that? Or is the issue more about processing power?
This was strangely motivating we all start off stumbling but over time we learn and grow I’m glad these two stickmen can now box
Would be interesting to see several ai bots put in a incredibly hard puzzle game where they need to cooperate to win.IT also will have this physics engine ofc and would be hilarious.
As a boxer for 8 years, this brought a tear to my eye
Actually, I would on the contrary be surprised if the researchers stopped at 130M steps. The reason was mentioned: the agents didn't even had a clue that they could get higher rewards by touching the opponent. So it was mainly a matter of random to start doing so, but until that it would be quite unreasonable to quit. But as always, it was still amazing to watch how did the imitate boxing :)
I'd like to see now a simulation of one bot that is after 1B steps against the bot that is after only, say, 100M steps. Would that improve learning speed of the second bot?
The one with the intentional ragdolling to trick the opponent was hilarious and clever.
I would love to see a video where you teach the ai different fighting styles and then had them fight other styles to see how they would adapt over time
That would be a great training tool for MMA professionals preparing for a fight. Not sure how hard it would be to incorporate grappling and submissions though.
we should try to make a simulation where movement costs them energy to see if they would avoid too many small and fast movements
Now I want to see an AI version of Robot Wars. Well described combatant rules, unlimited training. Last Bot standing wins.
@Femi Mark You wouldn't see the best though. You'd only see the best of the people that can afford it. Inevitably if more people can take part new ideas come in that even with only 200 hours, would beat the 3000 hour you'd have ended up with at the beginning. It's just they'd never have had an opportunity to showcase that idea.
How bout non sim training. You have to control ai spar partner to help it learn. I play toribash which let's you record moves as you perform them. It's fun but hard. Moves at designated frame rated. It's free on steam. If curious please check out let me know what you think.
@Zlysium this is where sponsorship becomes important in sports
The idea of such a game has been floating in my mind since about early 00's. Such game will be pretty niche and I don't believe it can be profitable yet. But as an entertainment, I'd love to have such game. But on the contrary, there are already plenty of games where players don't actually play, but instead program their bots, spending 95% of time for it, and only 5% of simply watching them playing. So, I believe it's not a huge deal to spend that same 95% of time on computing, given there is a simple and robust interface for that, so the players don't actually need to be experts in RL or DL.
This sounds like how you create an AI villain in a Marvel comic or something. 😳
It would be really neat if you could program each AI model with a different boxing style.
This man has just taught an AI how to beat the daylights out of something, so beautiful
1:13 In fighting games, we call this "conditioning" the opponent. We train the opponent to react to an action in a certain way, and then we punish them for reacting in that predictable way.
Now I want to see AI-controlled Toribash, with the AI particularly motivated to dismember its opponent.
Should’ve taught Tyron Woodley some of this
@Bryan And all punches before that are irrelevant
YYYYOOOOO! LMAO
Tyron only trained 130M steps before boxing Jake Paul
3 months later and it’s more true than ever 💀💀
@AguyfromHawaii Loooooool!!! 4:10
idk whether to be amazed or terrified lol. I understand the beauty of this but consequences seem inevitable.
I would be stoked to see video games AIs have systems to truly learn player behavior and other ai behavior that would be incredible
Yo! I think I've accidentally done the Adversarial thing before in Smash Bros. I forgot the stage and characters, and it was only for the first second of the match but I always did something that could be normalized, AKA something that I could buffer to do pretty much frame perfect almost every time. And I realized my AI opponent was also doing the same thing every time.
Just out of interest. Can you train 2 different fighters totally seperate. Then show the 2 fighters going at it? I wouldn't mind watching 2 ai who have 2 different training camps go head to head. Be a pretty interesting fight
I’m brand new to this stuff. Got very interested after training Power Platform’s object detector to find building components for my work. Does anyone have any suggestions on the best places to learn more about A. I.?
It looked to me like the red AI was placing its "head" as close to the ground as possible as quickly as possible, and the blue AI would follow that movement with its own "head" which unbalanced it. Why was it doing that? I don't know.
Were the AI's programmed to only be able to perceive what the other was doing via the "head" like a real human, so the blue would tilt its "head" forward to attend to the red?
The rear hand/power hand should offer an increased reward (just like a real cross offers increased power and damage if it lands) over the jab hand to stop it from becoming a stiff jab stalemate every exchange, and having fighters with slightly different dimensions as many others have said would also be a good change
This was very valuable to the personal research I'm doing on learning fir myself. Thank you, you did the experiment I was thinking of for. If I write a paper I hope to mention you in it.
0:45 I would be extremely impressed if a human thought of that but an AI thinking of something that clever is on a whole different level of impressive!
"Everyone has an algorithm 'till they get punched in the mouth." - AI Tyson
Mike TAIson
You sir, made my day. May you live a long and happy life.
You mean Ai Tyson
"Everyone has a plan until they get their ear bitten off" Mike Tyson
My man im dying
Is there any program that us normal people can run to watch this develop ourselves? I love watching this kind of stuff, and would love to experiment more with it in person!
We need to use these clumsy physics more in games. This is hilarious!
I believe many modern sports are going to learn from AI in the future. The learning curve AI has is way worse than from a human but its more compact, meaning it can train the same amount of time we take in a decade, in a few hours or maybe days. Its not like AI is so much more clever than humans.
That being said, i believe the AI will come up with strategies that pros will use in the future. Techniques that we didnt think of yet. It could be a huge stepping stone in terms of evolving modern sports. Cant wait to see it getting more advanced, to a level where pros are looking in awe.
Simulations are fine and good. But they're heavily irrelevant outside of testing stages. The real test will be how well the AI can adapt to a physical body in the outer world. Can't wait to see that.
I wonder how the AI would change if they were graded on knocking out the opponent instead of only touching them? Would it make them more aggressive or would they remain defensive?
I'd really love to see the fencing one, being a fencer myself it got me hype af
This could have tremendous interests for martial arts and fighting sports in general, to demonstrate which moves would be the most usefull and pratical if there's no teachers to teach you biased stuff and you only have an enormous amount of time practicing duelling.
All the old and possibly biaised tradition would be gone, and only the most practical things would remain.
it would be really cool if there were an array of different fighters with different proportions to see which ones beat which and how they adapt
The legs should have a separate ai from the torso/arms. The "puncher ai" tells the "walker ai" what it needs and the walker tries to achieve it. Also the puncher knows it's position the legs provide. The human brain can do this, unconscious actions take a small processing power because they aren't conscious. The conscious mind can take control, but doesn't need to.
i wonder the application of the video learning side of this. like can you use this to translate a video of a person moving or performing an action into a reusable 3d animation without the need of special suits or cameras? by combining the 3d physics system of how a body would behave and the single perspective camera i would imagine the AI would be able to clean up any missing perspectives ?
There was a game that was visually similar to this one, I don't remember the name, I only remember that they were stickmans and they fought, you more or less controlled their joints and only directed the blow, the physics of the game was very "ragdoll"
If anyone is familiar with what I am talking about and maybe they know the game I want to get to, I will be very, really, very grateful.
They can be confused with "Sumotori Dreams" but I'm not talking about that game.
Sorry if it is not well written, I speak Spanish.
wow, I wonder if the physics could modeled well enough that new boxing strategies (or strategies of other things) could be gleamed from such simulatations
Excellent video. Really interesting to see the different levels of improvement and the "hypnotism".
This is the second video I have seen from Two Minute Papers. Excellent cutting edge content. Well done.
I am a recreational boxer and a professional scientist. Model success can sometimes be limited by the 'game rules' hardwired into the core of the setup. For example, here we see the AI stick to lunging jabs at the end of training. The jab is the most efficient punch in boxing, so its understandable the AI learned this too, but solely relying on the jab doesn't really represent actual boxing. The reason for this is due to the AI being rewarded simply for touching the opponent without any consideration of power/speed of impact. Similarly, the while the AI successfully learned head movement, they are not blocking. This may be due to game rules penalizing the opponent's hand touching your hand (would discourage parries) or inappropriately high penalty for hits to the arms (would discourage hard shell blocking).
They should make them ambi, they only hit with the right hand which makes it look awkward.
also because they use the same AI for both fighters, they behave almost exactly the same which means they try to hit eachother on the exact same moment.
I'm no AI programmer by any means, but wouldn't it be better to alter parameters for one of the fighters and than combine both learning experiences of both fighters into one AI?
wongwu - 2021-08-29
"I fear not the AI that has trained in 10 billion simulations once. But I fear the AI that has trained in one simulation 10 billion times." - Bruce Lee probably
Raven Spades - 2022-09-08
@Warren Arnold oh yes i read that chapter
Vampire Hunter A - 2022-09-13
Bruce Lee IA
BoomPerson - 2022-10-24
It should be the other way around
Sanic07 - 2022-10-28
So you fear a one trick pony?
tropickman - 2022-10-29
This is basically TRIAL & ERROR to the maximum degree...