Why so few robot armys?

[SAMAS]

Don't forget; tanks are better. :v

In many settings it would be either irrelevant or change the focus of the story. Drama takes precedence over pet tech nonsense.

Great, now I got a mental image of an army of 2' robot tanks.

[jollyreaper]

i just recently finished reading the robots series by Isaac Asimov, as well as 2 of the culture books. a question can to my mind: why through out the vast universe(s) of each popular sci fi series(star trek, star wars, babylon5,star gate) there are so few instances of civilization using a wholly robotic military? i couldn't think of a logical for this other then taboo, or the writer didn't think it wouldn't be cool with them. are there any other reasons for this phenomenon?

Same reason why we don't see drone starfighters -- you need people in the mix to keep it interesting. If there's going to be a robot army, there's got to at least be humans involved fighting against them. If it's robots on robots or if there are clones which are essentially as faceless and anonymous as robots, it's very hard to get anyone to care.

In the real world, I think that battlefield automations will be as significant a development as gunpowder, the internal combustion engine, aviation and radio communications. I think 50 years from now will seem completely bewildering to us.

[someone_else]

Same reason why we don't see drone starfighters

A cruise missile duel! (optionally with "IN SPAAACE!!!" at the end)
How can that be made more interesting than a WWI-like dogfight with fighters piloted by female pilots with boobs?

Great, now I got a mental image of an army of 2' robot tanks.

Well, infantry usually is more than a walking gun. The bots need arms and hands to manipulate objects and doors, and the ability (and weight) to fit into the average house without collapsing floors. Also climbing stairs is kinda important.
That's closer to a 2' mecha.

[Starglider]

Same reason why we don't see drone starfighters

A cruise missile duel! (optionally with "IN SPAAACE!!!" at the end)
How can that be made more interesting than a WWI-like dogfight with fighters piloted by female pilots with boobs?

The cruise missiles could be piloted by Cortana-style AIs with impossibly perfect holographic bossoms. In fact I think this has already been done in anime, kinda (final battle of Macross Plus).

[Starglider]

There are reasons to want that on the battlefield, because otherwise you need a really massive debugging program to catch each problem in the robot's behavior, because humans are bad at writing foolproof logical instructions for a machine to follow.

AI software of this type does not gain much from (current) formal correctness proving techniques. The 'really massive debugging program' consists of running billions of simulated trials. The challenge being making the simulations realistic and comprehensive enough to catch the vast majority of undesireable behavior.

You save an enormous amount of trouble if the robot can learn and react to changing circumstances

There is no sane reason to equip military robots with individual learning capabilities (excepting some very basic adaptive pattern recognition to support specific tactical behaviours). It is vastly more sensible to periodically download all the sensor logs for offline analysis, by teams of engineers and supercomputers back at the facility which designed the robots. Artificial intelligence systems operating across networks at the theatre or strategic level are another kettle of fish entirely.

Batman was, believe it or not, arguing that because software is inherently unreliable, the safety lockout would have to be hardware. My entire point was that the lockout would of necessity be some kind of computer running a program, including a ROM chip.

Well, you could run some of the safety protocols in a separate watchdog CPU that the main CPU doesn't have write access to. That might be a sensible design choice depending on exactly what the risks are, e.g. if hostile forces have any significant chance of hacking into and taking control of the unit you'd want safety protocols to be non-overridable even if the hostiles can completely overwrite the mission parameters.

It's rather unlilely you will be able to make something so complex as a robot's brain out of analog components

Yes but there's no reason to, digital systems with appropriate comms isolation are just as good, at least assuming the attackers are other humans.

You save an enormous amount of trouble if the robot can learn and react to changing circumstances, but you also create the risk that the robot will malfunction in ways that start to look a lot like "rebel."

Which goes against the main mantra of militray "RELIABILITY ABOVE ALL".

Agree. I think most militaries would prefer to tolerate a higher attrition rate of equipment due to dumb/conservative robots than a higher incidence of friendly fire and misinterpreted orders.

If the software is unable to achieve a decent level of certainty, it will probably ask mommy just like soldiers do. "mommy" being their command officer.

Software can do this better than humans can actually, even if the commanding officer is tied up, it can at least find the first free operator at US Army Call Centre ShockAndAwe, flash up an enhanced image from its visual feed on their screen with a reticule around the hostile and a zoomed-in window showing their suspected weapon, and then pop up a dialog box 'Is this really a hostile? [ YES ] [ NO ]'. UAVs already have a basic version of this.

That's why stuff is tested. By "tested" I mean it is thrown in the hands of brave people that try to break it, and if after any problems they found are corrected, it reaches full production.
Alpha testing, beta testing and then final release.

Yes, that is relatively safe and reliable for robots up to about the chimp level of intelligence. I mean, you will probably still get the odd bit of bizarre emergent behavior in the field, but not worse than the standard occassional random SNAFUs you get with humans. And frankly chimp level intelligence is probably good enough for the things militaries are really keen to use robots for.

And a last clarification. I'm not against using self-improving softwares to create the program that will then be installed on the robot soldier.

This is the dangerous bit, much moreso than the robot soldiers themselves. In the future anyway, currently general systems for this are so hopelessly limited and unreliable that they're useless for this sort of specific robotic development task. People certainly use genetic algorithms, neural networks etc but for point solutions within specific human-defined task domains.

[Ford Prefect]

The cruise missiles could be piloted by Cortana-style AIs with impossibly perfect holographic bossoms. In fact I think this has already been done in anime, kinda (final battle of Macross Plus).

The X9 Ghost was totally autonomous, Sharon Apple simply asked it to go and kill Isamu and Guld.

[Simon_Jester]

There are reasons to want that on the battlefield, because otherwise you need a really massive debugging program to catch each problem in the robot's behavior, because humans are bad at writing foolproof logical instructions for a machine to follow.

AI software of this type does not gain much from (current) formal correctness proving techniques. The 'really massive debugging program' consists of running billions of simulated trials. The challenge being making the simulations realistic and comprehensive enough to catch the vast majority of undesireable behavior.

I'm inclined to agree.

See, I'm not qualified to say how you debug such a robot. I'm just pretty sure it's going to be a hell of a lot of work to do so.

You save an enormous amount of trouble if the robot can learn and react to changing circumstances

There is no sane reason to equip military robots with individual learning capabilities (excepting some very basic adaptive pattern recognition to support specific tactical behaviours). It is vastly more sensible to periodically download all the sensor logs for offline analysis, by teams of engineers and supercomputers back at the facility which designed the robots. Artificial intelligence systems operating across networks at the theatre or strategic level are another kettle of fish entirely.

A fair point.

Originally, I was trying to argue against the notion of an army of fully intelligent robots with "hardwired" provisions against rebellion externally imposed, as opposed to provisions that were integral to the robot's control software. I think I got spun round a bit.

It's rather unlilely you will be able to make something so complex as a robot's brain out of analog components

Yes but there's no reason to, digital systems with appropriate comms isolation are just as good, at least assuming the attackers are other humans.

I agree with this; Batman did not.

[jollyreaper]

The real question we cannot answer is whether sentient AI can be programmed or if it would have freewill like humans, our only example to work from.

We cannot program humans though we have all sort of tricks to compel obedience. Religion is incredibly powerful in that regard.

A predator drone can be programmed. Something like the original terminator was presented as simply a very advanced automated weapon. A guided missile homes on a target based on programming and the terminator is no different.

If we are talking about an AI with selfdirecrion like r2d2 in the previous example, I agre that programmed obedience seems unlikely. A restraining bolt could act like an electronic dog collar locking wheels if he moves to a restricted area but I don't see how it could block forbidden thoughts.

It would be interesting if AI minds would need to have taboos taught to them like we humans need to be taught incest, cannibalism, and furries are all taboo. The AI is taught obeying humans is holy writ, thou shalt not harm humans but it is something they deal with as guilt and conscience.

[Starglider]

The real question we cannot answer is whether sentient AI can be programmed

If nothing else, neurological evidence is overwhelming that the human brain is an information processing device amenable to artificial emulation.

or if it would have freewill like humans

'Free will' is a philosophical mirage, and should be ignored entirely in favor of much more concrete concepts when discussing cognitive engineering.

We cannot program humans though we have all sort of tricks to compel obedience.

We don't currently have direct write access to the neural net. Given sufficiently advanced technology, we could indeed program humans, although the structure of brains would make it a pretty messy endeavour.

If we are talking about an AI with selfdirecrion like r2d2 in the previous example, I agre that programmed obedience seems unlikely.

You are making a blanket comparison to human goal systems that is, from an engineering point of view, completely unfounded. When you engineer a system from scratch you choose what to put in, not what to leave out. A competently engineered system will not include desires to do unwanted things. It is regrettable that there are currently a lot of incompetent would-be AI designers out there who really do want to copy the human mind without bothering to work out how it works first.

A restraining bolt could act like an electronic dog collar locking wheels if he moves to a restricted area but I don't see how it could block forbidden thoughts.

Do you understand what thoughts are, in any given AI design that we are trying to 'restrain'? If you don't, then you are not qualified to have an opinion on this topic. In reality being able to tell what 'thoughts' are occuring is pretty vital to the design and debugging of an AI system. It is true that some (foolish) people choose system designs that are hard to white-box analyse, and that this kind of 'adversarial' mechanism will inevitably fail in the face of transhuman intelligence. This does not render such a mechanism infeasible or irrelevant when talking about relatively simplistic military robots with sharply limited learning capability.

It would be interesting if AI minds would need to have taboos taught to them like we humans need to be taught incest, cannibalism, and furries are all taboo.

Personally I would axiomatically define furries (in an AGI system) in as an example of superior transhuman aesthetics.

The AI is taught obeying humans is holy writ, thou shalt not harm humans but it is something they deal with as guilt and conscience.

Mixing AI and religion is a spectacularly bad idea, although trying to artificially and reliably replicate 'guilt' is close behind. These are about the worst possible motiviational systems you could imagine, and make absolutely no sense to anyone actually building autonomous robots.

[Srelex]

As a note in regards to the whole 'authors need human characters' thing, the comic strip 'ABC Warriors' depicts robots as consituting national armies, with them as the central characters--although it does so by basically making them tall metallic humans, right down to suffering 'psychological side-effects' and needing morale boosts.

[jollyreaper]

Starglider, until we actually have a working example of an AI brain, it's all still speculation. It's like coming up with aliens -- the alien I come up with might not be the same as yours but so long as my concept of it follows a consistent logic, doesn't break the laws of physics, and doesn't feature any gaping logic holes then whether or not you like it is a question of aesthetics. Now if I told you it was a jellyfish creature floating in the atmosphere of a gas giant and I also told you it developed spaceships, you would rightly ask "How can a floating jellyfish in a cloud build a spaceship?" If i don't have a really creative answer to that, there's a prime logic hole right there.

It's possible that thinking minds could be programmed but it's also possible that they might not. Seeing as we have no AI brains, you can make your assumptions that I can't disprove and I can make my own assumptions. I'm not sure who of us is closer to the truth, we might both be equally far away.

[Simon_Jester]

Thing is, Starglider almost certainly knows a lot more about AIs than you because he works on them for a living; I'd be inclined to trust him over you.

[Starglider]

Starglider, until we actually have a working example of an AI brain, it's all still speculation.

The entire field of AI would disagree with you. It's not as if robots magically pop out of the ether, fully formed. They exist because we design and build them. Researchers have and are investigating various strategies to build more sophisicated AIs. Many of these approaches aren't yet technologically feasible, but we have gained enough understanding to be able to make useful predictions about the characteristics of a fully developed system.

It's like coming up with aliens

No, it is not. These are not evolved creatures, they are designed and build by humans with specific goals in mind.

the alien I come up with might not be the same as yours but so long as my concept of it follows a consistent logic, doesn't break the laws of physics, and doesn't feature any gaping logic holes then whether or not you like it is a question of aesthetics

If you genuinely care about plausibility, this is dead wrong. Someone with a PhD in biochemistry will almost certainly come up with vastly more plausible concepts for alien creatures than either or us. It is extremely foolish to say 'the current state of AI is insufficient to predict what near-human robots will be like' without first having a very good understanding of the current state of AI. Admittedly this is a topic drift from fiction to futurism and real military robot development. If you only care about suspension of disbelief in your outright fantasies, then fine.

Seeing as we have no AI brains, you can make your assumptions that I can't disprove and I can make my own assumptions. I'm not sure who of us is closer to the truth, we might both be equally far away.

I am quite happy to explain the detailed reasoning behind the statements I have made, quoting references as necessary. It is admittedly a long inferential chain from the behavior of current prototypes to near-human AI, but no more so than using 20th century engineering knowledge to design interstellar space probes, to reference one quite technically impressive bit of hard futurism. You pulling assumptions out of your ass based on 'oh, I'll assume robots are like humans, because then I don't have to think too hard' is not comparable.

[Vendetta]

As a note in regards to the whole 'authors need human characters' thing, the comic strip 'ABC Warriors' depicts robots as consituting national armies, with them as the central characters--although it does so by basically making them tall metallic humans, right down to suffering 'psychological side-effects' and needing morale boosts.

Authors don't need human characters, they need characters. The advantage of a robot army is that you can mass produce it to spec, which is the thing which precludes having individual characterisation for the members of that robot army (otherwise there's effort involved in individual production, not truly mass produced, so they're not really what the thread was talking about).

[Chaotic Neutral]

Even if an AI needed to be made like a human brain, how difficult would putting in a "Rebellion = Massive Pain" mechanism be?

[Samuel]

About as long as it takes for the AI to erase that code because it interferes with its primary goals?

Authors don't need human characters, they need characters. The advantage of a robot army is that you can mass produce it to spec, which is the thing which precludes having individual characterisation for the members of that robot army (otherwise there's effort involved in individual production, not truly mass produced, so they're not really what the thread was talking about).

You can have characters even if they are massed produced- just have minor random variation among the troops to prevent them from being predictable or changes with experience.

[Chaotic Neutral]

About as long as it takes for the AI to erase that code because it interferes with its primary goals?

You mean like how when you break your leg, you rewrite you brain to ignore the pain signals?

[jollyreaper]

I am quite happy to explain the detailed reasoning behind the statements I have made, quoting references as necessary. It is admittedly a long inferential chain from the behavior of current prototypes to near-human AI, but no more so than using 20th century engineering knowledge to design interstellar space probes, to reference one quite technically impressive bit of hard futurism. You pulling assumptions out of your ass based on 'oh, I'll assume robots are like humans, because then I don't have to think too hard' is not comparable.

Please do. My understanding of the field is that we've managed to come up with fancy programs that can perform tasks used to require human input and we would have sworn an intelligence was required to deal with we now see is not the case. We keep redefining upwards what we're talking about when we mean AI and often have to ask if we're even using the same definitions when talking about these words. It's the same sort of misunderstanding of difficulties that saw us imagining clanking robots that could walk and talk to us and understand questions but would huff and puff dramatically to perform arithmetic calculations. The writers who came up with these robots did not appreciate that thousands of calculations per second would be required for the very act of walking, let alone the conversation bits.

As for my alien comment above, yes an expert in biochemistry will likely come up with an alien whose life processes are strange, exotic, and plausible, the same way a cultural anthropologist will likely create a more interesting alien society than I could or a linguist would devise a satisfyingly rich and varied alien language.

You are correct that we can make some decent guesses as to what near-future spaceships would be like and the engineers have the advantage there, assuming everyone agrees to stick to hard SF speculation and avoid woo-woo clarktech magic ships. At the same time, though, there are plenty of divisions between the hard SF advocates who have compelling ideas for how the future will work that are contradictory and there's room for dispute. There's always the classic question of "What makes more sense tactically, swarms of little ships with moderate firepower or fewer large ships with more firepower?" Completely depends on the technology at hand. Historically we've seen the advantage swing back and forth from dreadgnats to dreadnoughts.

[someone_else]

And a last clarification. I'm not against using self-improving softwares to create the program that will then be installed on the robot soldier.

This is the dangerous bit, much moreso than the robot soldiers themselves. In the future anyway, currently general systems for this are so hopelessly limited and unreliable that they're useless for this sort of specific robotic development task. People certainly use genetic algorithms, neural networks etc but for point solutions within specific human-defined task domains.

Why is this dangerous?
Anyway, I was thinking of low-level tasks (yeah, didn't said that in the quoted post), like learning how to use right arm/hand, left arm/hand, image recognition, and so on. The actual "superior brain functions" (i.e. the program that controls the robot's actions), will probably be easier to code than to have it learn.

I am quite happy to explain the detailed reasoning behind the statements I have made

Please do. I like to learn how things work.

We cannot program humans

Sure, there are some limitations (you need to do it while raising childrens from cradle to adulthood in a controlled environment) but that's pretty effective. On adults is usually a waste of time.
The main problem is that humans can learn, so can "deprogram" themselves to a certain extent if the situation changes (thus fucking up terribly whatever you told them).
Thus, to have effective "human drones" you need to keep them for most of the time in the abovementioned controlled environment.
Or, if you are truly good, you will be able to teach each individual to recreate his own little "controlled environment" wherever he lives.

Anyone will see the parallel to cults that aim to create secluded communities.

Even if an AI needed to be made like a human brain, how difficult would putting in a "Rebellion = Massive Pain" mechanism be?

"Rebellion" is a pretty complex behaviour, you cannot code it as an automatic reflex, since you will need another brain to recognize what is "rebellion" and what is not.

The best you can do is piggyback the "if you don't do what is your duty, you will feel bad, if you do your duty instead, you will feel good" that is already part of the human brain's works.
But this has the big problem of relying on a definiton of "duty" that, assuming the thing works like a human brain, can be changed through learning, misinterpreted, and/or mistaken, just as normal with us pink humies.

If you manage to codify the "correct" definition of "duty" and prevent your AI from changing it by whatever handwaving means you prefer (restraining bolts for example), and assuming you didn't leave planet-wide loopholes like the Three Laws, you will achieve what you want.
An AI incapable of doing what the creators think are bad things.

to reference one quite technically impressive bit of hard futurism. (Daedalus project)

Some high-quality images of that, also the site has lots of other cool images about a few other realistic spacecrafts

[Connor MacLeod]

It occurs to me that some, like Red and Shroomy, hit on some important points which boil down to "depends on" - as in, "depends on the semantics you use", "depends on the technology you have", "depends on how your culture/government/universe is structured," etc. I mean you can't even generalize that "robot armies are rare because people care about characters" and it is quite possible to have robot characters that are a product of mass-production (Star Wars and WALL-E come to mind.) Its all in HOW it happens.

If a person must have a "reason" for so few robot armies, the only possible answer is an out of universe one: Most authors, directors. scriptrwiters, and audiences in general probably neither care nor think too deeply about it, or give much thought to the consequences. In universe... you can pretty much contrive any number of explanations that are plausible (if not logical or rational.) as I think the thread demonstrates. Its probably along the same lines of why people often don't wonder why the sci fi weapons always fire slow moving glowy bolts, or wonder at why a handgun (or any sort of firearm or cannon or whatnot) in the hero's hand in any movie doesn't quite behave like they would in real life because they're firing blanks. Or the appearance of explosions, for that matter.

[guest]

Interesting debate, I think that the issue of AI rebellion is one of considerable interest. While many think that an artificial army rebelling is merely a hollywood myth, it must be noted that even at this point in time with the limited usage of AI controlled military equipment, there have been instances of either Al Queda or The Taliban(forgot which) using about a thousand dollars worth of computer hardware and about eighty dollars worth of software to hack into UAV's worth hundreds of thousands of dollars. While they didn't manage to "hijack" them, they did manage to gain access to the data feed coming out of the aircraft. If you introduce more complicated and intricate war machines to the battlefield it becomes increasingly likely that they could be hacked and even possibly hijacked.

There is also the issue that data corruption, or a program that gets changed to have a 1 put in a place where a 0 was supposed to be could result in your killbots turning and going after friendlies. The more autonomous you make the AI, the more complicated its programming has to be, the more circumstances the programmers have to account for, and the more likely they get that they forget one and a circumstance arises which will cause the AI to see an enemy where it should be friendly.

Then you've got the problem that the more complex and advanced the AI gets, the closer it gets to being completely sentient, which allows it to ask itself the question "why am I even doing this?"

The obvious solution is to hardwire the AI's to never want to rebel, much as humans are "hardwired" to reproduce. It isn't just something that we like to do, but it is something that we need to do. I mean a man works double shifts at his work to save up thirty grand, he loooooves the 1967 ford mustang, he fantasizes about it all the time, and wishes night and day to have one. In stead of taking the money and buying said car, he in stead starts spending it on that hot girl who cuts his hair every other week. He knows that she is a ditz, and he knows that she will probably dump him after a couple months, but he really wants to get her into bed. It doesn't matter that the car would give him countless hours of enjoyment for years to come, while the girl will give him good sex a few times for only a few months. He would rather start spending the money on the girl to get into bed with her.

And the drive to reproduce doesn't figure into it, because he knows that he can donate his sperm into a sperm bank where it will go on to fertilize some woman and create some kids. He could do that and castrate himself, and then he could go get the car and get the far more logical choice and, but he doesn't. There is about as much chance that the guy will voluntarily castrate himself, or stop thinking with his little head, because he is hard wired that way.

Much that way, we can make an AI controlled killbot about as unlikely to remove a control chip voluntarily as you or I would cut off a limb.

[Sarevok]

There is also the issue that data corruption, or a program that gets changed to have a 1 put in a place where a 0 was supposed to be could result in your killbots turning and going after friendlies. The more autonomous you make the AI, the more complicated its programming has to be, the more circumstances the programmers have to account for, and the more likely they get that they forget one and a circumstance arises which will cause the AI to see an enemy where it should be friendly.

The right number bits being corrupted to identify friendlies as enemies is very unlikely to happen on it's own. A wild array of things can go wrong when data is corrupted in a computer system. But the likely hood of something so specific like this happening on its own without malicious intent like a hacking attempt is impossibly low. In real world techniques like submitting corrupted data and buffer over attacks are common exploits. But they require a malicious minded intelligent individual to design and execute the attack.

[Starglider]

And a last clarification. I'm not against using self-improving softwares to create the program that will then be installed on the robot soldier.

This is the dangerous bit, much moreso than the robot soldiers themselves.

Why is this dangerous?

Because self-enhancing software is inherently unstable and it's very difficult to design reliable solutions for keeping its goals in line with the designer's intentions.

Anyway, I was thinking of low-level tasks (yeah, didn't said that in the quoted post), like learning how to use right arm/hand, left arm/hand, image recognition, and so on. The actual "superior brain functions" (i.e. the program that controls the robot's actions), will probably be easier to code than to have it learn.

Many research robots use this kind of hybrid approach, in that software engineers write the bulk of the software, but narrow machine learning (e.g. neural nets or support vector machines) is used to auto-generate some of the core functions. However software for most practical robots, AFAIK including all flying UAVs and the Big Dog prototypes, are just coded completely by hand.

I am quite happy to explain the detailed reasoning behind the statements I have made

Please do. I like to learn how things work.

I am on holiday in Sweden at the moment, you will have to wait until I get back.

Even if an AI needed to be made like a human brain, how difficult would putting in a "Rebellion = Massive Pain" mechanism be?

"Rebellion" is a pretty complex behaviour, you cannot code it as an automatic reflex, since you will need another brain to recognize what is "rebellion" and what is not.[/quote]

Exactly. And to make that work reliably it has to be more intelligent than the first brain, and then you have the problem of making the 'control' brain reliable. Each layer adds more complexity and more hidden failure modes. It is a diverging regress that cannot possibly work. Fortunately this problem does not get serious (at least in individual, mostly-isolated systems) until you closely approach human-level intelligence.

[someone_else]

The obvious solution is to hardwire the AI's to never want to rebel, much as humans are "hardwired" to reproduce.

Remeber that even in that field there is considerable variation and "screwups". Sodomy (non-reproduction oriented sex) and masturbation are kinda prevalent wherever you care to look. You also have feticists, homosexuals, transgenders, and whatever where sex is mostly for fun. (this is not to be interpreted as homophobic stuff, thank you )

And guess what? Animals do it too.

The main point is that the level of reliability that is acceptable for evolution is not anywhere near acceptable in an army.

Thus, without consistent handwaving, you cannot have a "totally reliable" AI made with a brain architecture.

The only way to have a "human-brain-like" AI is to remove the ability to learn new stuff.
It will be able to remeber, but not to learn. On mental tasks at least.

[sirocco]

Well when you look at all that has been said: a reliable AI with sufficient military technology would give an enormous asset on the battlefield. And that's enough to motivate some governments to require an international ban on full robotic armies (under the pretense that having to risk human lives would make people more responsible and less prone to MAD)

However as India, Pakistan, North Korea and potentially Iran showed it with nuclear technology, an international ban can be sometimes worthless.