@Karl with a K ah that doesn't matter friendo.. all you need is a gun attached to a quadcopter and the ability to hack the power grid, both of which have been possible for ages. if things ever get overwhelming just remind yourself "none of this matters anyways". it sincerely helps for me :)
I guess this one was just about moving around and swinging the sword in a convincing way. Have them fighting would require them to learn a sword style or even train them to invent and master their own techniques.
Imagine using this system in a gladiator simulation game where the gladiators are randomly generated with different movement and combat "styles" with a skill level that changes how smooth and refined the gladiator is. Seeing millions of unique potential combinations battle it out in various arenas with various environments in duels and team battles, different weapons and an ai that makes use of the environment well could be a next level video game for sure.
Chinese "Fake Martial Arts" masters do that since thousands of years :) But... could it not be used to find the really perfect martial art? That would be interesting.
Like Ray Kurtzweil recently said, “In the next 10 years, we will see 175 years”. This isn’t in relation to AI progress, we’re just going to shove 175 years into 10.
Just imagine if we could do this with humans? One year of Study and Research in one day, if something like this was possible, studying the whole universe would be easy.
At the rate they were going, they would train for 3652.425 years in ten years. Most likely more than that, since the hardware and software would likely improve during that time-frame.
the virtual space in which the AI operates will be very different and simplistic compared to the BD robot which functions in the real world. If the Virtual space can be made more realistic, your prediction may well be correct.
and so basically with the advancement of ai and assuming the robot can interact with the physical world as it would a simulated one, you essentially created artificial life.. ?
Player animations has always been the most immersion-breaking part of 3rd person games for me, and I’m really hopeful we will see massive improvements in big games soon!
@@MEATHEADBooYAI’m talking about janky transitions in between animations, like if you’ve ever tried wiggling back and forth quickly or jumping and landing in a run
@@arthurlefevre7706 even the best robots that we have today can't move that fluidly , its just not possible for them to move like we do , our muscles , joints and tendons are not even remotely comparable to any robot out there in terms of ease of movement , hopefully tesla bot does prove this wrong . We could have remote controlled robot battles for fun in the future then lol
that would be amazing, with some limits ofcourse. we would have to slow down learning drastically and also limit the amount of learning the enemies could do. otherwise the game would be unplayable within a few weeks because the enemies know every single possible action someone could take and be unstoppable :D
I've had a similar idea couple of years ago... enemies in certain area share the same AI which learns after each fight with a player to provide more challenge over the course of months or years
I see them clipping the swords through their skulls sometimes. If the researchers added damage zones to the heads and other parts, I bet the AI would learn to be more careful how they swung those weapons.
I don't know, maybe it's just an artifact from the visualization. The stick figure in the simulation is smaller in general and has also a smaller head.
A cool next step would be training the AI further, and then including the armor in the sims and giving it weight and collision. This would make the AI have to change its motion depending on how much armor it’s wearing and that would be the phenomenal to see purely from an investigative stand point.
This would be really cool for open world games. Imagine your character starts with a set basic movements, through playing it develops it’s own unique style of combat based of your movements and style. Mixed with multiplayer to create a truly unique combat style, where everyone’s movement and combat is different and optimized to their play style.
This sounds extremely fun and painfully hard to program. Especially since you'd have to consider what all the player controls and what the ai controls. Not to mention all characters would need at least a base level of training to ensure they can move otherwise you'd be sitting around for hours waiting. Although seeing an ai adapt to different weapons you pick up and armour could be interesting. Especially once you start factoring in stats like strength and agility. Seeing an ai struggle to walk in heavy armour, but adapting as time goes on or as you put more points into strength. Maybe agility is the fastest an ai could move a muscle meaning high levels of agility will result in fast dodging and being extremely hard to hit as the ai adapts to the skill boosts. Also the skill points wouldn't be as instant as other games since the ai would have to adapt to the changed variables before it can get the most out of it. All in all. I really want to get my hands on whatever system they are using to train these ais and learn how to code with them.
I wish, they'd put a kind of scanning or vision element into the network, so it can evade the boxes from the researchers and after 5 more years, we'll see generated matrix style fight scenes.
yeah I bet that's the next step. I suspect we'll see this stuff ramp up until it's just a "human animation" character, that can do practically anything. then once there is an open source, we won't ever need to animate characters again, because the 'actor' will just act and react accordingly. that's the end point of this; being able to direct these guys with spoken words, just like directors with real actors.
@@westingtyler1 I think the end point is at least doing that with a direct neural interface, not spoken words. But probably something beyond even that...
@@westingtyler1 Google's general purpose deep mind may be able to be trained much quicker than these and may even produce a better result in half the real world time
I love how, with this video especially, you add humor to these paper demonstrations! It truly makes the field of AI research seem more fun and approachable.
It'd be interesting to see them model things like realistic reaction times and muscle strength/fatigue, so the characters don't have perfect recoveries when they get knocked over.
yeah, the most immersion breaking thing was that they were too *good* at getting up from falls or the ground. its great if you were building a robot to have it never really fall, but for a game avatar like these its gotta go down and stay down from a big hit or it wont look human at all.
@@ensrceler your point about building a robot is really interesting, cuz it shows theoretically how agile combat robots that don't feel pain or surprise could be. the speed and smoothness with which they recovered is super impressive
This would be amazing if you could put physically simulated characters like this into VR games like Bonelab. They would make some of the most incredible and dynamic enemies in video games
I think it would make an interesting game where the ai enemies learn how the player moves and how the player fights and throughout the game they learn to imitate and counter you as the game progresses and the final boss is basically like fighting against yourself.
I notice it is using flat terrain with the same variables every time, I wonder how well it generalized for different terrain, gravity and character proportions.
Late comment, but any sloped or non-straight terrains, as long as they’re sufficiently smooth (not too jagged) are always able to be approximated by a flat plane in small patches (look up smooth manifolds). What developers could probably do is just piece together the whole terrain using a bunch of small little flat portions for which the model knows how to behave. This might be the easiest way to do it. Other things like character proportions/gravity, etc are likely all dependent on the actual physics engine of the game and the animators. You can’t micromanage the types of things AI does, you just hope it learns a general pattern of appropriate, realistic behaviors to “stimuli” (in quotes because there’s no real stimuli behind the scenes, it’s just a shit ton of matrices and calculus formulas that are constantly throwing out numbers).
This would easily become one of the most intimidating foes you could ever face in a video game Those high-energy recoveries are one of the most startling things to see them do honestly I wonder if it would be possible to link what actions they're tasked to perform with a player's controls (So that the AI acts as a sort of in-between for the player and their character, giving the character life-like movements while the player gives those movements purpose)
Does anyone know how soon this will be used in games? 😅 This is really amazing! Great work from NVIDIA, can't wait to see this technology in the next games I play! Thanks Two Minute Papers for covering this! 🙌
@Martin Kryl You're getting it wrong here. Currently there is a switch happening. We had the x86 processors for decades and now we are slowly switching to ARM. With those chips we will have way more power and way less energy consumption. Future prediction is: Small powerful phone always with you and at home or whatever you plug it into a dock with extra computing and more energy. So you will always have your data with you and can still experience games or whatever at maximum potential if you have the proper hardware.
@Martin Kryl i don't think the AAA game category will die, there are plenty every year, but i feel that ambitious projects are less frequent than before because most developers focus more on established IPs, and want to make their game available for as many ppl as possible (rare next gen only games)
@@pneumonoultramicroscopicsi4065 Wtf are you talking about have you seen what sony lately released and are still working on. Triple AAA games industry is bigger than the cheap ass movies that release these days.
Those recoveries were very impressive. They clearly have a lot more strength and stiffness than humans, but the time it took them to respond to the disturbance and correct their position is just.. wow
Imagine if you could simulate battles with this, giving units different equipment and pitching 2 teams against each other where the ai is able to fight for itself instead of having it based off of a percentage like the total war games
5:50 The from scratch simulation has so much unused potential. Imagine if you used an animation like that for Halo's Flood parasite combatants. So creepy. Just needs the right model, and environment.
Yes, but only if they are in the right place at the right time which is probably pretty unlikely. Otherwise here they make me cry laughing, which is fine, too, I guess. :)
It'd be great for a "realistic" version of a robot takeover! They wouldn't make the same mistakes as humans, but rather learn to move as efficiently as possible!
@@4P5MC I think the issue is they didn't calculate the impact force, penetration , recovery time and all sort of data that make a swing better than the creepy walk. Still can't stop laughing
2:27 that one in the left corner spinning around, when he dodged and did the shield bash it actually looked so realistic like it was an actual person in the suit doing it. Crazy.
AI training physics characters are my favorite thing to watch. In the future it will be used by robots to move around in the real world, but for now its pretty cool for videogames. For the next paper on this I want to see gameplay implementation or I am not holding on to my papers.
@@nickskywalker2568 True. There is also a possibility that AI like this will be used by governments. In the future, if robots become cheap enough to mass produce, technology like this can even be used in wars.
at 4:26 he even tried to kick away the incoming box. not sure if it was intentional or if we was doing the move anyway regardless of the box but it looked cool.
@@adamcsillag6058 They'll still have one major limitation, they can't tell what is an enemy and what isn't. So they'll likely to still have human commanders.
As I understood, there wasn't physics simulation of character body itself (muscles etc). You can see that looking on how they get up very quickly, almost flying.
This would be really cool to start incorporating, especially if you can add scripted Ai and modify their movements and skill sets. Once you know what you want you could set up the program at the start of whatever project you're on and then let it run until you're close to finishing it so maybe even like 2 irl years or so
It would be awesome to put a bunch of these agents in a fight, with them losing health after a hit, the damage scaling with how hard the hit was. It would be interesting to see how they use the shield, and if they develop their own fighting strategies. Maybe eventually this could be used to make a game like TABS.
This reminds me of messing around in Natural Motion's Endorphin back in the day. Incredible that this is now possible in real time while being more flexible as well. I really hope we get an interactive demo of some sort.
@@ItsWarFilms hm yes, i had endorphin, but naturalmotion also featured a lot their Euphoria ragdoll engine at the time that worked very well with GTA 4, that however required special access along with Morpheme th-cam.com/video/0Pm0Cvm0zdI/w-d-xo.html
I'd love to see the AI trained on reactions to hostiles with the characters having their own "fields of view" and reacting realistically to what is thrown at them. Rather than them just flipping back up onto their feet and getting thrown around again, I want to see them dodging, flipping, and blocking before they get hit. That would blow my mind and could lead to a world of better immersion in games.
I agree, and thats just the start. Imagine 3 more papers down the line when we could have this AI applied to real robots, making them absolute murder machines! What a time to have been alive!
4:15 this is my favorite section, because they hold up their sword like they're ready to go right after getting up. I know I'm anthropomorphising them, and it's probably just easier to gain balance that way, but it's still cool how ai can still have a warrior's spirit.
Can't wait to play video games with this model trained on stunt actors and martial arts masters. Just imagine what it'd look like on a game like Yakuza
There's is a game, "Absolver" I think was the name, that was all about fighting using martial arts styles. This could reduce the time needed to develop the styles and instead put that time toward fleshing out the game surrounding the combat (the issue Absolver had, too focused on the combat and not enough content to do it among).
@@rubyy.7374 The weakness of these games is the combat system which, despite of being really fun at start, tends to be repetitive and not very challenging. My point is by merging real martial arts with motion-captured supid heat actions, it could turn into a game which is both challenging and fun to play. Every knocked out punk should switch to a ragdoll algorithm tho, I fully agree.
Seems like a wonderful first application would be for generating background animations for NPCs. If nvidia made several trained sets available, or if large companies had the tools to train more, trading compute for art is a no brainer. I like the idea of every random training NPC at the first encampment in a game having an individualized and generated animation set. Or even a single very high quality one that didn't need to be hand animated or mocapped. Could be very good for indies someday
This would be an amazing way to provide player commands ie. movements/abilities/attacks. Let's say you're a low-level character, your movements are going to be standard albeit clumsy/slow. The more you train at things, the better and more advanced you get, and depending on the type of training, can make your character animations faster/stronger/weird depending on the training
Holy crap the output looks so realistic. They must have had all the anatomy and physics of the subject correct before they started the simulation. Like the stance and the way it moves its weight around when performing an attack looks as if it was mocapped from a pro athlete. And exactly how I was trained in Muay Thai IRL.
@@ShibbyRL yes the physics for the recovery doesnt look humane, when a human gets up the motion involves momentum balance and muscular tension throughout the body, here it looks like they can get up only using the strenght of only the part of their body that is in contact with the ground, negating momentum, which implies super human strenght
This just gave me a dope idea. Someone should make a game where each time you unlock an upgrade, your character gains a new move. Every core upgrade [make hard to get and only a few], the core upgrades dramatically improves chracter fluidity and movement with no change to the UI. That would be dope to start off with a clunky character but by the end game you have a super mobile, agile, powerhouse.
The steering gave me a lot of hype. I remember watching the E3 when the PS3 was being shown. When they were showing NBA 2k steering demo, I thought "wow the movement is so natural" and seeing this brings back those same thoughts.
Can't wait for games to have your character with a learning curve (based on XP) for their movements. And actually see a difference not only in hitpoint but in technic and movement also.
It's interesting to note how often they cross their legs as they step, this is something you work very hard to avoid in combat sports. I wonder if they would change when you introduce opponents or an artificial cost to falling.
Oh... duh... thats why all these sword fight videos look kind of goofy to me. I just thought the people were unathletic nerds, and probably alot are... but that explains a lot
For a lot of AI papers it would be great to mention what is the input of the AI, what it controls (its output), and what is its objective. We would better understand why some of these papers are such game changers.
Did you give it accurate weight? Like the weight of different pieces of armor and the weapon? It might make their movement less floaty and give it a real vibe, maybe change it as a whole.
Asking without having any knowledge about those A.I. programming, can this A.I. training be used to optimize real Robot locomotion like for example the Tesla Bot Optimus or the Hyundai / Boston Dynamics Robot Atlas ?
it would be really cool to have an open world RPG with different types of Ai, all learning to do combat with players. Over time, the developers can write in goals and ambitions for the Ai, making them good or evil. The idea would be to have a community based RPG where the story basically writes itself and is constantly changing based on super powerful Ai combatants that can fight with you, for you, betray you, or be loyal to you based onb how the community treats them. A neat twist might be to station these Ai in static locations for a little while they are in weaker phases. If a player kills them, they get a bit of loot. But if they die, the Ai gets the loot and can only be reclaimed once the player defeats the constantly evolving Ai combatant. Imagine going on a full scale raid to reclaim the loot that an Ai literally stole from you--with hundreds of players struggling against this mega powerful, world bending presence.
another exciting step on the road to realistic animated characters. in the paper abstract, this sentence stands out: "Our system also allows users to specify tasks through simple reward functions, and the skill embedding then enables the character to automatically synthesize complex and naturalistic strategies in order to achieve the task objectives." hopefully nvidia will soon allow ordinary users to specify those simple reward functions - perhaps including some less pugilistic moves!
imagine putting one of those little goal targets on the model's mouth and giving them a mug, we could FINALLY have NPCs that can actually drink things properly instead of clipping cups through their chins 😂 the possibility for other realistic little idle movements is really exciting though! I desperately want to get my hands on more of this new AI tech and play with it, I hope it will have good interfaces that are accessible to people that don't code much
@@Bo-kq8tn i think there is a reasonable chance of good interfaces. most things have a python API and a lot of 3D people know python. you may one day be able to able to instruct the AI in who drinks what without having to put targets on things. same will go for much else. then we shall be able to direct the avatars as if they were actors. i am hoping that we shall be able to use actors and AI together and increase speed of production, reduce cost and thus allow for much better, more interesting, more intelligent and more unusual stories - which the market otherwise filters out of existence on the grounds of cost for too small an audience. it is even possible - and certainly desirable - that there would be more work for actors not less, and more interesting work too. perhaps fewer megstars, but maybe that would be no bad thing for ordinary, talented but generally unknown actors and performers.
So does this mean in the future games will have combatants that act independently and adapt to the conditions their put in? This seems like it would really benefit VR where the players actions are basically unlimited,
Probably not much better, AI training tends to level off after a while, and they're already matching the dataset moves so there's no more improvement to go really.
Would love a strategy game based on this. Player doesn't take direct control, but acts like an agent at an explorer's guild or something to influence their bots to grow up into fine warriors with a literal textbox.
This could be really good for background characters in movies, and anime when you have a lot of fighters in the scene. For example, this could be used to animate an army, making the animated soldiers perform differing movements.
its interesting, but its still only transitioning between a few animations, so while it looks really really good, its motions are limited. Id really like to see some combat where it picks moves based on the opponents moves, like an actual fight. But seeing nvidia choose combat robots is a bit scary, are they planning to make police bots based on these with shields and swords?
Damn, I wish to see video games that use this in the future. Actual dynamic movements and fights with no preset motions, but actual procedurally generated movements, attacks and defences ^^
the closest I've seen to real looking ai movement was The Last of Us II, and with that they spent years honing the movement system. that is until now. this looks incredible! can't wait to see this be implemented in games and movies. slowly getting rid of key framing.
Does anyone realize how insane this is going to be for VR games? Imagine a game where you are just a greek soldier and you get dropped into a battle with 1000 of these guys.
I would love to see this used to create the perfect perhuman fighting style. Imagine taking simulation of ill the worlds fighting styles and having the AI’s practice and mix them together with a neural network and creating a new fighting styles. It might even create new stances that maximizes the human body. The simulation with score points passed on how it harms it's opponents and how quickly it get there. Maybe it could have the models use biological human models to get accurate damage and joint restrictions
this wouldnt work because these simulations are pretty basics , there are millions of factors in real life that arent being taken account for in these simulations
@@chakibboumaraf3127 totally understandable I used to wrestle and many moves are very situational and these same moves wouldn't work in something like boxing. I was mostly thinking of having different body types and heights randomized and starting from different distances and postures and completing specific fighting styles this would give the system a baseline of what works best in what situation from there using that data the system could determine what martial art are best used when. this is where you would start the serious number crunching when making various changes in the style and are moves are used where. it would be as interesting as a complete martial arts made from scratch by AI but it would allow a more fluid transition to best known styles and when you use them no matter the body type.
Clearly this would be a great way to scale the combat difficulty in video games. instead of pumping up the amount of health/dmg an enemy does..you scale the enemy to have more "experience" so to speak. Put it on super easy..and its like "1 week of training" difficulty Put it on very hard and it's like "25 years of training" difficulty. You plug the AI into the type of enemy..and they learn to use the locomotive parts to create a "smart" A.I. for video game npcs.
I want to see them battle each other! It would be interesting to see the different techniques the ai could come up with. Also it would be realy interesting o see how they line up against real life sword techniques.
It took a bit for it to click in my head, but this is a infant combat AI If you simulate it in a realistic environment those AI could be put into a robotic chassis similar to their simulated one and they would continue to learn Badass but also kinda scary considering 10 days makes them somewhat competent, imagine one learning for a whole year while having a physical body for half of it
I have long wondered what motivates recursive learning in AI systems. Is it positive or negative reinforcement? I can't help but wonder if those running away were whipped and tortured into returning to the battlefield... Because those that do return don't seem to fight they just seem to get right into the heart of the battle and try not to fight anybody.
Thank you so much for delivering research in a way that is understandable to the general public. Also, I love your inspirational messages interspersed throughout your video.
Well explained. Thanks 👍
@Karl with a K ah that doesn't matter friendo.. all you need is a gun attached to a quadcopter and the ability to hack the power grid, both of which have been possible for ages.
if things ever get overwhelming just remind yourself "none of this matters anyways".
it sincerely helps for me :)
More ai is what I like
So nobody gonna talk about how we are training the robots fighting skills so they can overtake human society some day?
robot's*
Well explained. Thanks 👍
I'm a little upset they didnt have them fight each other. I would have loved a large scale fight.
You have to put them in a ballroom to make them do that.
maybe they kept that one for the next week lol
I guess this one was just about moving around and swinging the sword in a convincing way.
Have them fighting would require them to learn a sword style or even train them to invent and master their own techniques.
@@FelanLP Which could be the next step
@@Galaxia53 Which hopefully is the next step
Imagine using this system in a gladiator simulation game where the gladiators are randomly generated with different movement and combat "styles" with a skill level that changes how smooth and refined the gladiator is. Seeing millions of unique potential combinations battle it out in various arenas with various environments in duels and team battles, different weapons and an ai that makes use of the environment well could be a next level video game for sure.
add a dash of gambling to that and i guarantee it will be big
Like TABS, but more customisable and harder to render
Your pc would explode like a c4
@@Smokedouttasian oh man, gambling on gladiators like the olden days.
Chinese "Fake Martial Arts" masters do that since thousands of years :)
But... could it not be used to find the really perfect martial art? That would be interesting.
Like Ray Kurtzweil recently said, “In the next 10 years, we will see 175 years”. This isn’t in relation to AI progress, we’re just going to shove 175 years into 10.
Well, I think it goes deeper than that, since 175 years of training on their system is only 5.75 months!
It sucks, everybody is unprepared. Many people might die in the process due to sync or distribution issues.
That's the best way yet to describe the confusion of the present world.
Just imagine if we could do this with humans? One year of Study and Research in one day, if something like this was possible, studying the whole universe would be easy.
At the rate they were going, they would train for 3652.425 years in ten years. Most likely more than that, since the hardware and software would likely improve during that time-frame.
So basically you just need to put this trained AI into the Atlus Boston Dynamics robot and the matrix is no longer a work of fiction. damn.
There are different variables to account for in a virtual simulation, compared to the real world, but that doesn't mean it's impossible.
the virtual space in which the AI operates will be very different and simplistic compared to the BD robot which functions in the real world. If the Virtual space can be made more realistic, your prediction may well be correct.
and so basically with the advancement of ai and assuming the robot can interact with the physical world as it would a simulated one, you essentially created artificial life.. ?
@@nickalveberg3364 yep... the last child of humanity more or less.
Combine this with chatgpt and we are screwed...
Player animations has always been the most immersion-breaking part of 3rd person games for me, and I’m really hopeful we will see massive improvements in big games soon!
even out of the gaming realm, imagine this training applied to biped robots in real life.
Its not that hard now days, as the suits for doing animations have come a long way.
@@MEATHEADBooYAI’m talking about janky transitions in between animations, like if you’ve ever tried wiggling back and forth quickly or jumping and landing in a run
Please fix Skyrim. Thanks.
@@arthurlefevre7706 even the best robots that we have today can't move that fluidly , its just not possible for them to move like we do , our muscles , joints and tendons are not even remotely comparable to any robot out there in terms of ease of movement , hopefully tesla bot does prove this wrong . We could have remote controlled robot battles for fun in the future then lol
I'd love to see learning enemies in a mmo type videogame so both players and the game would adapt to have a constant challenge
that would be amazing, with some limits ofcourse. we would have to slow down learning drastically and also limit the amount of learning the enemies could do. otherwise the game would be unplayable within a few weeks because the enemies know every single possible action someone could take and be unstoppable :D
@@Aiphares I would guess it works if the character which is played has a lot of moves to do and the AI would forgot some once in a while.
Imagine fighting one in a blade and sorcery vr game
I've had a similar idea couple of years ago... enemies in certain area share the same AI which learns after each fight with a player to provide more challenge over the course of months or years
I remember oblivion doing something similar. Enemies would level up gears and stats as the player leveled up. Was such an annoying feature haha
I see them clipping the swords through their skulls sometimes. If the researchers added damage zones to the heads and other parts, I bet the AI would learn to be more careful how they swung those weapons.
It depends on how they feel about those other ai’s. 😁
I don't know, maybe it's just an artifact from the visualization. The stick figure in the simulation is smaller in general and has also a smaller head.
A cool next step would be training the AI further, and then including the armor in the sims and giving it weight and collision. This would make the AI have to change its motion depending on how much armor it’s wearing and that would be the phenomenal to see purely from an investigative stand point.
This would be really cool for open world games. Imagine your character starts with a set basic movements, through playing it develops it’s own unique style of combat based of your movements and style. Mixed with multiplayer to create a truly unique combat style, where everyone’s movement and combat is different and optimized to their play style.
This sounds extremely fun and painfully hard to program. Especially since you'd have to consider what all the player controls and what the ai controls. Not to mention all characters would need at least a base level of training to ensure they can move otherwise you'd be sitting around for hours waiting.
Although seeing an ai adapt to different weapons you pick up and armour could be interesting.
Especially once you start factoring in stats like strength and agility. Seeing an ai struggle to walk in heavy armour, but adapting as time goes on or as you put more points into strength. Maybe agility is the fastest an ai could move a muscle meaning high levels of agility will result in fast dodging and being extremely hard to hit as the ai adapts to the skill boosts.
Also the skill points wouldn't be as instant as other games since the ai would have to adapt to the changed variables before it can get the most out of it.
All in all. I really want to get my hands on whatever system they are using to train these ais and learn how to code with them.
I wish, they'd put a kind of scanning or vision element into the network, so it can evade the boxes from the researchers and after 5 more years, we'll see generated matrix style fight scenes.
yeah I bet that's the next step. I suspect we'll see this stuff ramp up until it's just a "human animation" character, that can do practically anything. then once there is an open source, we won't ever need to animate characters again, because the 'actor' will just act and react accordingly. that's the end point of this; being able to direct these guys with spoken words, just like directors with real actors.
@@westingtyler1 And then I will edit The Mask into the Matrix movie.
They won't do it. I mean... We won't see it. Military development isn't something for public, you know...
@@westingtyler1 I think the end point is at least doing that with a direct neural interface, not spoken words. But probably something beyond even that...
@@westingtyler1 Google's general purpose deep mind may be able to be trained much quicker than these and may even produce a better result in half the real world time
I love how, with this video especially, you add humor to these paper demonstrations! It truly makes the field of AI research seem more fun and approachable.
It'd be interesting to see them model things like realistic reaction times and muscle strength/fatigue, so the characters don't have perfect recoveries when they get knocked over.
yeah, the most immersion breaking thing was that they were too *good* at getting up from falls or the ground. its great if you were building a robot to have it never really fall, but for a game avatar like these its gotta go down and stay down from a big hit or it wont look human at all.
@@ensrceler your point about building a robot is really interesting, cuz it shows theoretically how agile combat robots that don't feel pain or surprise could be. the speed and smoothness with which they recovered is super impressive
This would be amazing if you could put physically simulated characters like this into VR games like Bonelab. They would make some of the most incredible and dynamic enemies in video games
This idea underated as heck, and it would stop ford from getting up when im grabbing him
The way the sword works with latent space is really interesting. Can't wait to see where this research ends up after a few more papers!
I hope they implement awareness for self-collision soon because the time of unwanted intersections must end somewhere.
I think it would make an interesting game where the ai enemies learn how the player moves and how the player fights and throughout the game they learn to imitate and counter you as the game progresses and the final boss is basically like fighting against yourself.
That might be fun, but sometimes it's more fun to fight against someone better than yourself.
I notice it is using flat terrain with the same variables every time, I wonder how well it generalized for different terrain, gravity and character proportions.
just a few papers down the line
They probably need to stumble down a very gentle incline for a few million iterations before they get the hang of that
Late comment, but any sloped or non-straight terrains, as long as they’re sufficiently smooth (not too jagged) are always able to be approximated by a flat plane in small patches (look up smooth manifolds). What developers could probably do is just piece together the whole terrain using a bunch of small little flat portions for which the model knows how to behave. This might be the easiest way to do it. Other things like character proportions/gravity, etc are likely all dependent on the actual physics engine of the game and the animators. You can’t micromanage the types of things AI does, you just hope it learns a general pattern of appropriate, realistic behaviors to “stimuli” (in quotes because there’s no real stimuli behind the scenes, it’s just a shit ton of matrices and calculus formulas that are constantly throwing out numbers).
This would easily become one of the most intimidating foes you could ever face in a video game
Those high-energy recoveries are one of the most startling things to see them do honestly
I wonder if it would be possible to link what actions they're tasked to perform with a player's controls
(So that the AI acts as a sort of in-between for the player and their character, giving the character life-like movements while the player gives those movements purpose)
Does anyone know how soon this will be used in games? 😅
This is really amazing! Great work from NVIDIA, can't wait to see this technology in the next games I play! Thanks Two Minute Papers for covering this! 🙌
@Martin Kryl Good point. But I do not blame the kids these days when looking at the GPU prices over the last year(s).
@Martin Kryl You're getting it wrong here.
Currently there is a switch happening. We had the x86 processors for decades and now we are slowly switching to ARM.
With those chips we will have way more power and way less energy consumption.
Future prediction is: Small powerful phone always with you and at home or whatever you plug it into a dock with extra computing and more energy. So you will always have your data with you and can still experience games or whatever at maximum potential if you have the proper hardware.
@Martin Kryl i don't think the AAA game category will die, there are plenty every year, but i feel that ambitious projects are less frequent than before because most developers focus more on established IPs, and want to make their game available for as many ppl as possible (rare next gen only games)
@@pneumonoultramicroscopicsi4065 Wtf are you talking about have you seen what sony lately released and are still working on. Triple AAA games industry is bigger than the cheap ass movies that release these days.
@@mrX666-s9p Sony is yet to release a next gen only game
5:55 the cursed movement on the left 😂
Those recoveries were very impressive. They clearly have a lot more strength and stiffness than humans, but the time it took them to respond to the disturbance and correct their position is just.. wow
It's the result of an optimized algorithm and 100% dedication to their training ;)
Imagine if you could simulate battles with this, giving units different equipment and pitching 2 teams against each other where the ai is able to fight for itself instead of having it based off of a percentage like the total war games
Totally accurate battle simulator already does this
Looking forward to see them in non-levelled terrain
😂
Another 10 years😉
Already done check a company called embark studios and what they have done with there in game AI
@@vesaruntti I see what you did there
That's cool. Learning to chain together movements and making it look realistic. I wonder how it would "feel" in a game
5:50
The from scratch simulation has so much unused potential.
Imagine if you used an animation like that for Halo's Flood parasite combatants.
So creepy. Just needs the right model, and environment.
yo actually true
Yes, but only if they are in the right place at the right time which is probably pretty unlikely. Otherwise here they make me cry laughing, which is fine, too, I guess. :)
It'd be great for a "realistic" version of a robot takeover! They wouldn't make the same mistakes as humans, but rather learn to move as efficiently as possible!
@@4P5MC Humans already moving pretty efficient. So efficient they beat at long distance every mammal on planet.
@@4P5MC I think the issue is they didn't calculate the impact force, penetration , recovery time and all sort of data that make a swing better than the creepy walk. Still can't stop laughing
2:27 that one in the left corner spinning around, when he dodged and did the shield bash it actually looked so realistic like it was an actual person in the suit doing it. Crazy.
AI training physics characters are my favorite thing to watch. In the future it will be used by robots to move around in the real world, but for now its pretty cool for videogames.
For the next paper on this I want to see gameplay implementation or I am not holding on to my papers.
Let's just hope we get clean governments in the meantime...
@@nickskywalker2568 True. There is also a possibility that AI like this will be used by governments. In the future, if robots become cheap enough to mass produce, technology like this can even be used in wars.
@@ET-yc4wb It's a 99% chances possibility
@@ET-yc4wb only wars? it can be used by ANY law enforcement agency AND private companies with enough money, but I'm sure it will be ok...
@@lingred975 ofc not only wars. I just used wars as a worst case scenario.
Brilliant! - Any chance you can do a similar simulation to make them play the bagpipes?
at 4:26 he even tried to kick away the incoming box. not sure if it was intentional or if we was doing the move anyway regardless of the box but it looked cool.
It's terrifying to see them practicing combat in a life-like simulation.
Yes imagine they will get a robot body and actual weapons, with say like 1000 years of training. Terrifying.
@@adamcsillag6058 if it does get to that point they’ll probably simulate several millennia worth of time before putting the ai in its body
@@adamcsillag6058 They'll still have one major limitation, they can't tell what is an enemy and what isn't. So they'll likely to still have human commanders.
@@robbieaulia6462 Oh that's not a limitation rather an advantage. They will kill anyone.
As I understood, there wasn't physics simulation of character body itself (muscles etc). You can see that looking on how they get up very quickly, almost flying.
This would be really cool to start incorporating, especially if you can add scripted Ai and modify their movements and skill sets. Once you know what you want you could set up the program at the start of whatever project you're on and then let it run until you're close to finishing it so maybe even like 2 irl years or so
I'm afraid training an ai for 2 irl years would violate the first law of papers.
It would be awesome to put a bunch of these agents in a fight, with them losing health after a hit, the damage scaling with how hard the hit was. It would be interesting to see how they use the shield, and if they develop their own fighting strategies.
Maybe eventually this could be used to make a game like TABS.
I want see these guys fighting a boss with Skull reaper or gleam eyes texture from sword art online
im glad someone knows tabs
TABS that is actually a TABS
to much data is bad for ai training, doubt it would yield the quality of results you are hoping for
@@crwelch12 oh
This channel is the sole reason that I’m so interested in AI and ML. Thank you for all your hard work!
This reminds me of messing around in Natural Motion's Endorphin back in the day.
Incredible that this is now possible in real time while being more flexible as well. I really hope we get an interactive demo of some sort.
I always wanted to use euphoria but now the technology is so outdated XD
@@2k7u the program is still really fun to play with actually.
@@ItsWarFilms hm yes, i had endorphin, but naturalmotion also featured a lot their Euphoria ragdoll engine at the time that worked very well with GTA 4, that however required special access along with Morpheme th-cam.com/video/0Pm0Cvm0zdI/w-d-xo.html
Careful what you wish for...
@5:51 No dude that is legit AF! Imagine going in a ring with a competent fighter and he attacks you like that. xD
I'd love to see the AI trained on reactions to hostiles with the characters having their own "fields of view" and reacting realistically to what is thrown at them. Rather than them just flipping back up onto their feet and getting thrown around again, I want to see them dodging, flipping, and blocking before they get hit. That would blow my mind and could lead to a world of better immersion in games.
I agree, and thats just the start. Imagine 3 more papers down the line when we could have this AI applied to real robots, making them absolute murder machines! What a time to have been alive!
@@flinkstiff Sound like suicide with extra steps XD
@@trevorhook5677 would be if you made them sapient
@@shiggermetimbers sentient?
4:15 this is my favorite section, because they hold up their sword like they're ready to go right after getting up. I know I'm anthropomorphising them, and it's probably just easier to gain balance that way, but it's still cool how ai can still have a warrior's spirit.
Probably noy, they are trained to move like real warrior.
So this move was probably in the dataset.
Can't wait to play video games with this model trained on stunt actors and martial arts masters.
Just imagine what it'd look like on a game like Yakuza
There's is a game, "Absolver" I think was the name, that was all about fighting using martial arts styles. This could reduce the time needed to develop the styles and instead put that time toward fleshing out the game surrounding the combat (the issue Absolver had, too focused on the combat and not enough content to do it among).
I do hope they keep the funni ragdolling though.
@@rubyy.7374 The weakness of these games is the combat system which, despite of being really fun at start, tends to be repetitive and not very challenging.
My point is by merging real martial arts with motion-captured supid heat actions, it could turn into a game which is both challenging and fun to play.
Every knocked out punk should switch to a ragdoll algorithm tho, I fully agree.
I've gotta say I love how enthusiastic you are, it's contagious
I also liked the way they encouraged the AI to explore different movements in the latent space while training.
I hate how addicting this channel can be. I mean like, I've watched his videos for the past 3 hours and still aren't tired yet
Seems like a wonderful first application would be for generating background animations for NPCs. If nvidia made several trained sets available, or if large companies had the tools to train more, trading compute for art is a no brainer.
I like the idea of every random training NPC at the first encampment in a game having an individualized and generated animation set. Or even a single very high quality one that didn't need to be hand animated or mocapped. Could be very good for indies someday
it looks like it'd be amazing for Swords and Sorcery lol
Or crowd simulation in movies/tv
This would be an amazing way to provide player commands ie. movements/abilities/attacks. Let's say you're a low-level character, your movements are going to be standard albeit clumsy/slow. The more you train at things, the better and more advanced you get, and depending on the type of training, can make your character animations faster/stronger/weird depending on the training
Holy crap the output looks so realistic. They must have had all the anatomy and physics of the subject correct before they started the simulation. Like the stance and the way it moves its weight around when performing an attack looks as if it was mocapped from a pro athlete. And exactly how I was trained in Muay Thai IRL.
The physics looks like what’s been done years ago in euphoria engine in gta 4
Now everything will be Euphoria simulation for any game, + physical particles data for\of surfaces and materia. 😉
the falling and the get back up animation looks so cool. really feels like a trained warrior recovering from an unexpected blow.
for a scifi version yes, i feel like if the recovery wasnt as fast it would be truly realistic. But super cool none the less
@@ShibbyRL yes the physics for the recovery doesnt look humane, when a human gets up the motion involves momentum balance and muscular tension throughout the body, here it looks like they can get up only using the strenght of only the part of their body that is in contact with the ground, negating momentum, which implies super human strenght
@@vrpnblstr3441 considering the programm doesn't render millions of muscle fibres, but like 10 bones instead, yeah
@@vrpnblstr3441 for video game standards I don't think that matters
This just gave me a dope idea. Someone should make a game where each time you unlock an upgrade, your character gains a new move. Every core upgrade [make hard to get and only a few], the core upgrades dramatically improves chracter fluidity and movement with no change to the UI. That would be dope to start off with a clunky character but by the end game you have a super mobile, agile, powerhouse.
Can’t wait for the sort immersion this opens in interactive media. Being able to give voice commands to AI squads would be so amazing
The steering gave me a lot of hype. I remember watching the E3 when the PS3 was being shown. When they were showing NBA 2k steering demo, I thought "wow the movement is so natural" and seeing this brings back those same thoughts.
Can't wait for games to have your character with a learning curve (based on XP) for their movements.
And actually see a difference not only in hitpoint but in technic and movement also.
Apply the model to a humanoid robot with the same DoF. That would be crazy!
It's interesting to note how often they cross their legs as they step, this is something you work very hard to avoid in combat sports. I wonder if they would change when you introduce opponents or an artificial cost to falling.
Oh... duh... thats why all these sword fight videos look kind of goofy to me. I just thought the people were unathletic nerds, and probably alot are... but that explains a lot
@@michaelwerkov3438 It's not an easy concept but keep digging, you might eventually understand why it looks goofy and why everyone is so exited.
The ways they get up are insanely cool! Like the best break dance moves I've seen in my life
For a lot of AI papers it would be great to mention what is the input of the AI, what it controls (its output), and what is its objective. We would better understand why some of these papers are such game changers.
i believe the input is mocap data provided by reallusion
Did you give it accurate weight? Like the weight of different pieces of armor and the weapon? It might make their movement less floaty and give it a real vibe, maybe change it as a whole.
10 years into 10 days!??!? Imagine what Ai could do if 10 centuries was compressed into one year!?!?!😵💫😵💫
Imagine the speed of AI when quantum computers are being made 😫🥶
Asking without having any knowledge about those A.I. programming, can this A.I. training be used to optimize real Robot locomotion like for example the Tesla Bot Optimus or the Hyundai / Boston Dynamics Robot Atlas ?
I do not have the words to be able to fully express how absolutely amazing this paper is.
5:50 When the mosquito gets inside your helmet.
kek
I feel like this would be great in VR games, either as NPC's or player-controlled characters.
it would be really cool to have an open world RPG with different types of Ai, all learning to do combat with players. Over time, the developers can write in goals and ambitions for the Ai, making them good or evil. The idea would be to have a community based RPG where the story basically writes itself and is constantly changing based on super powerful Ai combatants that can fight with you, for you, betray you, or be loyal to you based onb how the community treats them.
A neat twist might be to station these Ai in static locations for a little while they are in weaker phases. If a player kills them, they get a bit of loot. But if they die, the Ai gets the loot and can only be reclaimed once the player defeats the constantly evolving Ai combatant. Imagine going on a full scale raid to reclaim the loot that an Ai literally stole from you--with hundreds of players struggling against this mega powerful, world bending presence.
another exciting step on the road to realistic animated characters. in the paper abstract, this sentence stands out: "Our system also allows users to specify tasks through simple reward functions, and the skill embedding then enables the character to automatically synthesize complex and naturalistic strategies in order to achieve the task objectives." hopefully nvidia will soon allow ordinary users to specify those simple reward functions - perhaps including some less pugilistic moves!
How did you comment 3 days ago, when this video was posted like 1 minute ago?
@@mackenziestorey620 channel member
imagine putting one of those little goal targets on the model's mouth and giving them a mug, we could FINALLY have NPCs that can actually drink things properly instead of clipping cups through their chins 😂 the possibility for other realistic little idle movements is really exciting though! I desperately want to get my hands on more of this new AI tech and play with it, I hope it will have good interfaces that are accessible to people that don't code much
@@Bo-kq8tn i think there is a reasonable chance of good interfaces. most things have a python API and a lot of 3D people know python. you may one day be able to able to instruct the AI in who drinks what without having to put targets on things. same will go for much else. then we shall be able to direct the avatars as if they were actors. i am hoping that we shall be able to use actors and AI together and increase speed of production, reduce cost and thus allow for much better, more interesting, more intelligent and more unusual stories - which the market otherwise filters out of existence on the grounds of cost for too small an audience. it is even possible - and certainly desirable - that there would be more work for actors not less, and more interesting work too. perhaps fewer megstars, but maybe that would be no bad thing for ordinary, talented but generally unknown actors and performers.
at 4:00 the second from the right in the bottom standing up is really cool!
So does this mean in the future games will have combatants that act independently and adapt to the conditions their put in?
This seems like it would really benefit VR where the players actions are basically unlimited,
Bugger all gamers.
They look like more modern versions of storm troopers. I love those illuminated shields and swords. They look so cool.
This is what I imagined the game Clumsy Ninja to be like before I actually played it
These virtual guys stand up with more style than most of the viewers 😭
I would love to see them train the data for 100 days (years) and see how the maneuvers advance.
Probably not much better, AI training tends to level off after a while, and they're already matching the dataset moves so there's no more improvement to go really.
@@theaveragepro1749 They have gone beyond the dataset, as all AI do. But I do agree that skills such as this do tend to cap at some point in time.
i like the guy in the back right corner, who just walks calmly, then squats.
Would love to see such movements in games such as Sifu and Absolver.
Would love a strategy game based on this. Player doesn't take direct control, but acts like an agent at an explorer's guild or something to influence their bots to grow up into fine warriors with a literal textbox.
This could be really good for background characters in movies, and anime when you have a lot of fighters in the scene. For example, this could be used to animate an army, making the animated soldiers perform differing movements.
“But they still look like constipated warriors” lmao dude they totally do. And now that association will forever be in my head
its interesting, but its still only transitioning between a few animations, so while it looks really really good, its motions are limited. Id really like to see some combat where it picks moves based on the opponents moves, like an actual fight.
But seeing nvidia choose combat robots is a bit scary, are they planning to make police bots based on these with shields and swords?
I wanna simulate the most graceful and deadly Iaido practitioner defending himself from different scenarios.
Its possible to use that AI in game project?
Get this comment higher so it is Answered
Damn, I wish to see video games that use this in the future. Actual dynamic movements and fights with no preset motions, but actual procedurally generated movements, attacks and defences ^^
the closest I've seen to real looking ai movement was The Last of Us II, and with that they spent years honing the movement system.
that is until now. this looks incredible! can't wait to see this be implemented in games and movies. slowly getting rid of key framing.
4:02 the one on the bottom left had possibly the coolest recovery of all time.
I just bought a RTX 3090 GPU. Looking forward to try these mind blowing things out :)
How much
@@DiselSun at least a dollar
fighting each other would require techniques(block an attack, attack on open, run away when sword was thrown by enemy and etc.) base on a situation.
Does anyone realize how insane this is going to be for VR games? Imagine a game where you are just a greek soldier and you get dropped into a battle with 1000 of these guys.
Enjoying your sense of humour man. Really made the intro funny! Cheers.
Can't wait to fight these after the AI uprising
NO WAY THIS IS NOT PRE LOADED ANIMATIONS
I would love to see this used to create the perfect perhuman fighting style. Imagine taking simulation of ill the worlds fighting styles and having the AI’s practice and mix them together with a neural network and creating a new fighting styles. It might even create new stances that maximizes the human body. The simulation with score points passed on how it harms it's opponents and how quickly it get there. Maybe it could have the models use biological human models to get accurate damage and joint restrictions
this wouldnt work because these simulations are pretty basics , there are millions of factors in real life that arent being taken account for in these simulations
@@chakibboumaraf3127 totally understandable I used to wrestle and many moves are very situational and these same moves wouldn't work in something like boxing. I was mostly thinking of having different body types and heights randomized and starting from different distances and postures and completing specific fighting styles this would give the system a baseline of what works best in what situation from there using that data the system could determine what martial art are best used when. this is where you would start the serious number crunching when making various changes in the style and are moves are used where. it would be as interesting as a complete martial arts made from scratch by AI but it would allow a more fluid transition to best known styles and when you use them no matter the body type.
Clearly this would be a great way to scale the combat difficulty in video games. instead of pumping up the amount of health/dmg an enemy does..you scale the enemy to have more "experience" so to speak.
Put it on super easy..and its like "1 week of training" difficulty
Put it on very hard and it's like "25 years of training" difficulty.
You plug the AI into the type of enemy..and they learn to use the locomotive parts to create a "smart" A.I. for video game npcs.
The thing is 10 years is a long time to train an AI.
1 day = 1 year for them
2:32 You didn’t watch the video
@@allstar4065 I did watch it. I meant in reality no company would train an AI for 10 years
I want to see them battle each other! It would be interesting to see the different techniques the ai could come up with. Also it would be realy interesting o see how they line up against real life sword techniques.
My papers were flying off the shelf and putting holes in my walls while watching this. I could NOT stop them
The 2 year "going to leave... in style" cracked me up!
Just had one of those really satisfying "well, than answered my question" situations.
the movement and weight look so fluid and realistic
It took a bit for it to click in my head, but this is a infant combat AI
If you simulate it in a realistic environment those AI could be put into a robotic chassis similar to their simulated one and they would continue to learn
Badass but also kinda scary considering 10 days makes them somewhat competent, imagine one learning for a whole year while having a physical body for half of it
Amazing explanation!!
Hopefully games will start using stuff like this and not pre-made animations. This is sorta looks like Euphoria but better.
"First lesson is to learn how to fall." Love it. Thanks!
I'm scared that this shows the potential of a well-trained robot soldier in reality as well as virtual reality
I have long wondered what motivates recursive learning in AI systems.
Is it positive or negative reinforcement?
I can't help but wonder if those running away were whipped and tortured into returning to the battlefield... Because those that do return don't seem to fight they just seem to get right into the heart of the battle and try not to fight anybody.
i have an idea for you : make thoses ia to start fighting each other to see how their technic evolve !
Thank you so much for delivering research in a way that is understandable to the general public. Also, I love your inspirational messages interspersed throughout your video.