Slashdot Mirror
On Game AI In The Uncanny Valley
An anonymous reader writes "Normally, the Uncanny Valley theory is used to critique graphical realism in games, but it also applies to AI. Therefore, designer David Hayward examines AI's Uncanny Valley over at Gamasutra, citing games from Valve's Half-Life 2 to the interactive drama Façade." From the article: 'There's a small minority of people who are consistently strange in particular ways... I don't mean to pick on them as a group; nearly all of us dip into such behavior sometimes, perhaps when we're upset, out of sorts, or drunk. Relative and variable as our social skills are, AI is nowhere near such a sophisticated level of interactive ability. It is, however, robotic. Monstrous and sometimes unintentionally comedic; the intersection of broken AI and spooky people is coming.'"
Elmo on fire
liqbase
One of the things which people don't mention about the uncanny valley, is that the valley moves. What seems uncannily human to our parents, was normal to us. What's uncanny to us, will look artificial to our kids.
As our technology improves to create better and better artificial representations, our ability to detect them does as well.
The ______ Agenda
You're in a game running along, there's someone beside you... Then you look at him, and he looks back, but there's nothing inside there...
Playing with bots and people, REALLY unpleasant.
I suspect that consultation with and evaluation by psychology departments may become relevant to game AI in the coming years, given that they're the most comprehensive resource in existence on human behavior
:)
I disagree with this. I think in the future, game programmers won't have to go as far as the psychology departments of their nearest schools. They'll just have to walk over to the nearest cubicle and talk to the animators working on the game. As game models have become more and more complex, companies are using more and more motion capture to capture action sequences, but animators (especially good ones) are trained to make a non-living 3d model imitate human behavior. There's over 50 years of research done for animation by animators on how to bring life to drawings and 3d models in motion - which is something that can be directly transferred over into programming terms as opposed to a research paper on a psychological disorder. An animator can tell you when to make a character blink in order for it to appear more realistic; a psychologist, not so much.
btw, IAAA (I am an animator) so I'm slightly biased
Don't trust a bull's horn, a doberman's tooth, a runaway horse or me.
The game the writer keeps talking about "Facade," invites the player to play multiple run-throughs. The writer then is creep-out by the "robotic" behavior that surfaces. Since without a time machine it is impossible to view humans in the same way, how do we know we would fair better. Bill Murry's character in the movie "Groundhog Day" seemed to develop a view of his fellow man not different than the view expressed by the writer of this article.
. . . do we want in a video game? Do we want game AI that prevents NPC's from running straight into a head shot or AI that causes us to pause out of guilt before making that head shot? Game AI usually refers to the ability of a game to employ varied strategy, adapt to player strategy, and generally be unpredictable without cheating.
AI/Robotics are barely at the cute level ... long way to go before functional in social settings.
t em
... "Vapor/Spoof-Ware" applies to software, hardware, content management, drugs, knowledge management ... many more things .... It all (marketeers) reminds me of the old snake-oil sales history of two centuries ago.
Game theory and Complex Systems mathematics expressed 3/4D visually will greatly reduce wait time for virtually natural (complex and non-repulsive) AI/Robotics and human collaboration/relationships. From the familiar to the revolting to acceptance we will go over the next (maybe) 30 years.
http://necsi.org/
http://www.complex-systems.com/
http://en.wikipedia.org/wiki/Complex_adaptive_sys
Presently the biggest problem to retail/business investment in technology is marketing spin-bullshit. To many things sound like, but in reality ain't shit, or at most is not as advertised
Damn, again I went a little off track, Sorry about that.
Unaccountable leaders are masters, and unrepresented people are slaves. How do US and EU fare?
I'd argue that you yourself, anonymous coward, are also merely a very deeply iterated switch statement, as am I, and everyone around us. Where in your thought process does anything other than brain hardware and past experience enter into the equation?
I used to carry a bottle of whiskey for snake bite. And two snakes. -Nefarious Wheel
my guess is, it started getting called 'intelligent software' when it started to have more 'if' 'ands' and 'buts', because this new level of complexity made previous software 'simple' by contrast.
I would assume that sophisticated game AI algorithms wouldn't use boolean logic as much as some kind of "fuzzy" logic. That is, it would take into account the number, strength and proximity of known enemies, adding up to some "threat level", compare that to its own "capability level" which includes things like health, available firepower, proximity of friendlies, etc., then applies a "personality factor" which makes that unit more or less agressive, etc. All that is fed into a function that decides the next move.
While humans have many advantages, when you're talking about combat, most soldiers react based on their training because there isn't a lot of time to think. Therefore it may be able to create AI that follows the same basic "twitch" style of gameplay as a human if you tried hard. The difference would come after the encounter when a human reflects on the game and adapts their gameplay, whereas the AI would act virtually the same every time.
"I have never let my schooling interfere with my education." - Mark Twain
>> [...] the intersection of broken AI and spooky people is coming.
Ever "talk" with Eliza?
That's a broken personality right there...
I don't know the meaning of the word 'don't' - J
The author is missing the point of the uncanny valley. The point of the uncanny valley is the dip. If you have a non-realistic pixar-like character (The Incredibles), you empathise with them. If you have a more realistic, but not good enough character, the character gets creepier (Polar Express). Then, when it gets really good, you can empathise with them again (Hugo Weaving, i.e. Agent Smith, in the Matrix Sequels).
The author basically says "AI is hard". But he doesn't make any real argument as to there being some "valley" where as the characters get less realistic they act more believably, and as they get more realistic, they also act more believably. A much more accurate title would be "AI is a steep hill".
I think we're a long way off from seeing computer controlled characters that aren't merely following scripts. AI in gaming is in its infancy and in many cases is non-existent. I see the term applied to numerous games and inevitably come away underwhelmed by AI that isn't any more sophisticated than what I was seeing in Pac Man. Well, I'll concede AI is more sophisticated nowadays given the more complex nature of games and the fact that they're inhabiting a 3-dimensional space nowadays. Still, that isn't saying much.
Ultimately, development of AI is going to be dictated by budget. The first question is, does AI really make a game more entertaining? In some cases it does not. So why invest time and effort into overly sophisticated AI? This becomes an issue especially when massive resources have already been dumped into graphics and the gaming engine itself. It seems most developers can't even be bothered with a more sophisticated NPC AI in RPGs.
I remember years ago those applications that would attempt to mimic a psychologist by asking all sorts of questions and attempting to attempting to appropriately respond to the user's answers. It was a bit goofy, but in my opinion far more sophisticated than the responses NPCs in most RPGs spit out. And honestly, beyond the basic engine I don't see it requiring much work beyond some additional writing.
Beyond that, it's small touches that can make an NPC look alive. As another poster mentioned a big part of it animation. It would be interesting to look over at an NPC, have that NPC notice this and turn it's head to look at you. All this requires a considerable amount of production, because a single life-like movement isn't going to suffice.
What I'd really like to see is a game with AI that learns, be it an FPS or ideally something else. Essentially a game has a character that the player encounters on a regular basis. That AI learns from the player and improves itself. Ultimately these characters can be shared online and be pitted against other players. Then maybe we could possibly have truly challenging AI as opposed to AI that merely "cheats" with perfect aim and unrealistically quick reaction times.
Either it's very challenging or I'm just no aware of what's being done out there but I'm surprised more game developers haven't explored the possibilities of AI.
The catch, of course, is that all this animation comes at a pretty huge cost in time. It takes teams of animators countless hours to generate all that animation.
I keep thinking that procedural animation is going to be the next big thing. Instead of rigid animations, we'll see rules governing the position of each limb and how they interact with the world and other characters. It's expensive in terms of CPU processing power, but it allows for a far more natural interaction with the environment. There's a movie of the new Indiana Jones game that shows this off nicely.
My dream, as a hobbyist game developer, is to have animation and text to voice systems get to the point where I can do a lot of the work myself. As it is, coordinating voice actors and model animation takes a huge chunk of time.
what you said is true, but you do not see past our ability to do things contrary to all of our 'experience'.
for example, i know if i touch a hot flame, it will burn me and it will hurt. i may not want to touch the flame, but i can decide to anyway, even against my own will/judgement/past experience and every other possible variable that can be taken into account. AI is almost the opposite. it's static. even if you were to throw in a "random" variable that ends up making the final decision between any number of choices, the fact remains that the final decision rests on the "random" variable that it is forced to use, so, its 'statically random'. there is no real choice in software. there is choice in human intelligence.
I would assume that sophisticated game AI algorithms wouldn't use boolean logic as much as some kind of "fuzzy" logic. That is, it would take into account the number, strength and proximity of known enemies, adding up to some "threat level", compare that to its own "capability level" which includes things like health, available firepower, proximity of friendlies, etc., then applies a "personality factor" which makes that unit more or less agressive, etc. All that is fed into a function that decides the next move.
While humans have many advantages, when you're talking about combat, most soldiers react based on their training because there isn't a lot of time to think. Therefore it may be able to create AI that follows the same basic "twitch" style of gameplay as a human if you tried hard. The difference would come after the encounter when a human reflects on the game and adapts their gameplay, whereas the AI would act virtually the same every time.
First and foremost, game AI as we know it has almost nothing to do with real AI for several simple reasons. First of all, the game would simply be too darn hard for the players if it used real AI modeling. If true AI approaches were used to create different base logic units, or "personalities" that had the ability to "learn" the environment and world, games quickly become too difficult to play and win.
It is not altogether difficult to create a Kohonen neural network and teach it how to react to certain situations. A rewards based learning and evolution environment will quickly be able to generate different personalities for use in a game. It is very easy in fact for game AI, as there are a finite number of inputs that need to be generated depending on how complex you want to go. Really there are probably only about dozen personalities that need to be made for any "shooter" games. You can always make more if you want to have more interesting interactions, but that leads me to the second reason why real AI is not used in games, it is NOT predictable.
You can not script it to do something. Each time you play, it will work a little differently. Complex interactions quickly occur with even the simplest of reward rule sets. You will be surprised when the game AI guards start killing the berzerker units becuase the guards learn that the guards around the berzerker units die faster becuase the "player" tends to use a rocket launcher on the berzerkers and guards located near the berzerkers tend to die as a result... You can start to see my point here, the game itself become un-predictible, which most story line games can not be, certain events have to occur, certain areas must be difficult to make it through, etc., etc., what good is it if the AI learns that it has no chance to fight the player and win, and thus decides to run away? Now some of that can be fixed by the reward system for the AI units, i.e. give a large reward for activily engaging the human player, but then you will run into cases where you will have units that snipe if there is a rule in place that rewards the AI unit for staying alive as there almost always should be (otherwise you just wind up with suiciders).
My point is that games and real AI will not work for many of the games we currently have and play. Some games which are not storyline driven, things like MMORPGs or open ended games may very well benefit from true AI by changing the monotony of the game itself. But it is just cheaper and more controlled in the minds of the publishers to not do this and hard code behaviour so they have a more consistant product.
We were all warned a long time ago that MS products sucked, remember the Magic 8 Ball said, "Outlook not so good"
How are you sure there's choice in human intelligence? Yes, you can 'decide' to touch fire anyway if you want. But why did you want to? To put out the fire? To make a point? All you've shown is that your program is more complex than "fire == pain"; that whatever is static in your program is at a higher level*. Granted, the static parts of game AIs are usually far less general, but it's a difference of amount, not kind.
* Though in some sense the only invariant in the human brain is physics, it's probably not necessary to simulate intelligence at such a low level to be convincing. You could go at least as high as neurons and their interactions, and possibly even to the level of subsystems like vision, motor coordination, etc. Much of the structure and function of your brain is static.
First, the Uncanny Valley is a hypthesis. Second, it does not state that realistic renderings look weird. Instead, it states that humans show more and more empathy as models become more and more realistic, but that this rule breaks if we approach a perfect model, and humans suddenly look like zombies.
This can be easily observed. The unrealistic humans in "the incredibles" seem much more human than the children from "Polar Express," even though Polar Express uses a much more realistic rendering style.
Wikipedia has a good article on the topic.
I think you mean a feedforward neural network instead of a Kohonen one. And even then you're overestimating how smart learning algorithms will be compared to engineered ones like state machines and production rules.
You're right about the predictability and scripting though. This is exactly to avoid the uncanny valley. Automatic AI keeps them dumb, but fun to play against/with and when they need to behave more like humans, a human inputs exactly what they should do or say.
I still call bull. I was one of those who quoted the Uncanny Valley left and right, but, sorry, I'm more and more convinced that it's so much bullshit it could fertilize a few acres. There is no one-dimension axis measuring likeness to human.
E.g., let's take two sets of models which were both in the "uncanny valley" if it exists. On one end you have the extremely detailed models of Final Fantasy: The Spirits Within, and on the other the models of EQ2 which a _lot_ of people described as "lifeless", "sterile" or other such euphemisms. I'm one of those. In fact I'd call them just disturbingly wrong, the kind where your subconscious keeps snapping out of suspension of disbelief screaming "that's not a tree!" and "that's not a human!" If an uncanny valley exists, then they're in the uncanny valley too, right?
Then something that falls in between those two points should be in the uncanny valley too, right?
Well, wrong. Oblivion for example was a lot easier to swallow than either, although detail-level-wise it's between the two.
That's called a Reductio ad absurdum proof, where assuming X leads to the false conclusion Y.
Furthermore, if you've actually read the Uncanny Valley theory, the examples used are blatantly bogus hand-waving. E.g., yes, a zombie is disturbing, but it's bogus to claim that it's purely for aesthetic or "how much it resembles humans" reasons. There's a whole bunch of cultural and emotional meaning tied to that, and claiming that it trips people's fears just because of the "uncanny valley" effect, is like claiming that you fear a car coming at you just because the headlights look sorta uncanny like eyes.
And again, you can do a reductio ad absurdum there. The Undead in WoW are the most disturbing visually, the characters in Spirits Within are uncanny valley too, so something in between should be in the Uncanny Valley too. Yet the Wow humans and elves are considered the races that look good.
In fact the zombies are the perfect counter-example all by themselves. If lowering realism moves something out of the uncanny valley (e.g., lower polycount characters in games are less disturbing than the characters in The Spirits Within), then it should do the same for zombies. It doesn't. They're still repulsive even at Quake 1 polycounts.
Or if slight imperfections like in Spirits Within, or details like teeth on a vampire, are what makes them disturbing via an Uncanny Valley effect... then how about pointy ears on elves? Shouldn't Legolas cause the same effect? Well, bummer, he doesn't.
Etc, etc, etc.
The Uncanny Valley is one of those things that makes sense only as long as you don't actually use your brains.
Yes, there are all sorts of ways in which being different or acting not-right can trip people's suspension of disbelief. That much is obvious. But there is no single dimension measuring it, and no single Uncanny Valley graph. There are thousands of factors which can be right or awfully wrong or somewhere in between, and thousands of fears, beliefs, expectations that can be tripped by it. You can't take the average and use it as the X axis for an uncanny valley graph, because even if the average is 99% right (hence the whole should be on the right side of the valley), one single detail (e.g., "omg, they're zombies") which can be disturbing on its own.
E.g., The Spirits Within wasn't just "a littel off", it had outright bad acting. That's what tripped people's suspension of disbelief. There was no uncanny valley effect, no overall being just a little off, it's just what you'd get with human actors acting badly.
It also overlooks the problem of expectations. The Spirits Within is wrong because you expect them to be human, Toy Story or Oblivion aren't because you expect them to be respectively toys or NPCs in a computer game. You have different sets of expectations for them.
Aesop's Fables (since they keep getting mentioned as Uncanny Valley effect exam
A polar bear is a cartesian bear after a coordinate transform.
D6 63 0D 70 89 81 BB 8E 7B 7C 5F 5D 54 EA AB 73
Exactly where do you see a false dillema? No, seriously, quotes about possible logical fallacies are good and fine, but you must first establish that a logic error has actually been commited.
The logic I'm proposing is along the lines of:
- _if_ there's such an axis and the curve looks like that (i.e., with a single dip below zero), then the "uncanny valley" zone is basically an interval
- if two points are part of the same interval, then a point in the middle is part of that interval too. (E.g., taking the interval [0.5..0.9], if 0.6 is in that interval and 0.8 is in that interval, then so is 0.7.)
- point X is in that interval (e.g., X = The Spirits Within.)
- point Y is in that interval (e.g., Y = EQ2)
- point Z is between X and Y (e.g., Z = Oblivion. It does have both better models and better animations than EQ2, yet far worse than The Spirits Within.)
- then point Z should be in that Uncanny Valley interval too
- worse yet, because of the shape of that curve, Z can't end up looking better/less-disturbing than either X or Y, if both X and Y are inside that dip below zero.
And yet that's false. Z actually looks a hell of a lot better, and trips people's suspension of disbelief a lot less.
If you can find a genuine logic error in the above, please do let me know.
And that, in a nutshell, is the problem with the whole Uncanny Valley _bullshit_. It's easy to hand-pick examples that confirm whatever you wish, and it's easy to hand-wave any single example as to why it should fall inside or outside of it by itself. It becomes hard when you start noticing that any two points you choose, you can pick a counter-example in the middle which doesn't act that way.
The whole Uncanny Valley bullshit is itself based on two fallacies, since you brought those up:
1. Biased Sample. You can pretend to prove anything if you're free to hand-pick only the examples that confirm it.
2. Not as much a logical as a mathematical fallacy: being fuzzy enough to be able to redefine the curve as to explain anything you want explained. Any single example you can choose, someone can (and will) handwave and massage its position or the shape of the curve to justify that it really fits their preconception. Since you can't put a clear number and say "zombies are 75.61% realistic", it just allows people to handwave where they should be placed to confirm an unproven preconception.
It's a pity that the Texas Sharpshooter Fallacy name is already taken, because it's pretty much the perfect illustration of what the proponents of the Uncanny Valley bullshit constantly do. Any single point you choose, they can just use a ton of sophistry and handwaving to move the point or the curve to explain it. It's literally painting the target around the bullet hole, instead of honestly testing a hypothesis.
3. Often using a Inconsistent Comparison or an Incomplete Comparison fallacy to create the illusion of that unidimensional scale where one doesn't exist. Every single example is judged by whatever different criterion fits the preconception, avoiding the aspects which would make it a lot harder to squeeze in a single variable.
Well, I'm attacking just that stupid handwaving, and especially number 2. I'm using two points to freakin' anchor that curve already and show that a third can't possibly fit it.
A polar bear is a cartesian bear after a coordinate transform.
When something does not appear human, its human traits stand out. When something does appear human, its inhuman traits stannd out. This is basically what the Uncannny Valley is about. Exactly what should be considered human and inhuman here probably depends on a thousands of factors. Perhaps being "done well" is part of it, but mostly I think more human appearance and behaviour raises the stanndards of craftmanship considerably. When you're using stick figures, it's not a problem if they don't move exactly like Tiger Woods. In fact, it might be uncannny if they did. If you've got spectacular, photorealistic graphics, everything that's not equally realistic gives you the feeling that something is not quite right.
I for one can say that Eliza actually improved my social life a lot. You just need to realize that you can do just that in a conversation too.
;)
Are you stuck in a conversation with a zealot ranting about how BSD is better than Linux, or why Gentoo sucks compared to the Gay Penguin distro? Have to listen to someone ranting about what subtle differences make Manowar the greatest band ever, and Metalica just a bunch of soulless sell-outs? Have to nod through as your GF goes on at great lengths as to what should some soap-opera character do, or such? Stuck with a co-worker ranting every single day, for two years straight, about doing the same things in the same fucking CounterStrike map you've never heard of?
Well, ok, so there's the easy "if you don't talk to me ever again, I will refrain from breaking both your legs, cutting your balls with a rusty saw, and shoving them so far up your ass you'll gag" way out, but bear with me anyway
The other way is to realize that you too can go through a whole conversation by just rephrasing what they've said.
Is that guy going on about how "emerge" sucks compared to the "bend-over-and-take-it" script in his favourite Gay Penguin distro? Never fear, just wait for the right moment to feed that bit right back at him. Sooner or later he'll give you some excuse to say, "Yeah, well, that's why emerge sucks compared to bend-over-and-take-it" or "well, duh, with bend-over-and-take-it it would have went much better" right back at him. Do it right, and he'll think he's found the perfect guy, who shares all his opinions to the letter.
Do remember though Eliza has the problem that it does it (A) immediately, and (B) quite unskilled at getting it to be a grammatically-correct sentence, much less one that makes sense. You can do better. For starters, don't feed them their own stuff back immediately. Throw a little delay in the feedback loop. Especially make sure that 10 seconds have elapsed. The short term memory buffer on most humans is 8 seconds, and it's good to have a bit of a margin, just in case.
A polar bear is a cartesian bear after a coordinate transform.
Few games try to have dynamic AI employed for textual NPC dialog, but Animal Crossing does ... sort of. I remember when I first got the game for the GCN, I was under the assumption that everything they said was dynamically plucked from every letter I wrote. Then after a few weeks I started to realize that it was just mad libs, and their responses to my letter were always generic.
I think I played Animal Crossing Wild World longer than the original one (but still not as long as the game expected me to play it, over a year? Pftt). But I played it long enough to tell that they hadn't really put that much effort in the AI. Each character had one of like six personality templates, and those templates were just madlibs where they would have blank spots that would be plugged in with human-written or quotes from letters or random items names. Meh.
If a game expects me to do but interact with AI for years at a time, then that AI damn well better be dynamic.
was extremely complex, how do you interact with it. Your input into a game is severely limited. The game knows nothing of you but your keypresses. You'd have to hook yourself up to your computer to get any kind of real interaction. Let it measure your heartrate, body temperature, where your gaze is and your whole face in general. Voice recognition with a good way to measure all the nuances of emotion would go a long way in itself. As it is, the most complex input you can give a game is text, and that conveys almost no emotion, while introducing several ambiguities. Until we can really jack into a computer and feed it more body language, it's still going to have a helluva hard time convincing us of interaction because those poor NPCs won't have but the slightest ideas about us.
Incidentally, Airport security seems to be approaching this level of biometric-response measurement, so if no one destroys the planet, maybe we'll see something.
Black and White was a terrible game but the part that sticks out is the avatar you get to create. You can use the carrot and stick method to train that beastie. When it misbehaves, you get to smack it around. The reaction of this 3D model to getting beaten made me feel ill in a way that your typical shooters did not. I mean, this is a cute little animated critter and the way it reacted to getting hit and the wimpering sounds, it was like watching a mother batter her child. I gave up on the game because it sucked but if it was good, I'm not sure how I would have handled that avatar.
In the first Aliens Vs. Predator game for the PC, the humans had some nice scripted reactions for when your alien or predator attacked. The marines could flip out, either spraying bullets at random or dropping their weapons and running. Civilians would flee as best they could. If you cornered them they would cower and crouch down, wimpering. It made the game feel a lot less like a fancy shooting gallery and more like something you were doing to real people. It didn't help that the gore effects were exceptionally vivid with dismemberments, heads getting stuck to walls with predator spears, etc. How will it feel when they get the characters looking and behaving as realistically as the good Disney theatrical animation? How long until it looks pretty much photo-realistic? I think a lot of gamers would be turned off if they were playing something like Doom except with the violence level as real as a soldiercam in Iraq. I think there could be a trend away from photorealism for precisely that sort of reason.
Kwisatz Haderach
Sell the spice to CHOAM
This Mahdi took Shaddam's Throne
If it's bad, it's not uncanny at all. It's just bad. The Uncannny Valley is what happens when it's extremely good. A previous article about this on slashdot explained that the same thing happens when you see a fresh corpse. It looks perfectly human, but something subtle, something really important. is missing.