More Human Than Human: The Rise of Robotics Technology

coming up on today's podcast the recent progress in humanoid robots is such that we may well be seeing them as the primary occupant in the kitchen but also a bit like AI itself it's not just an app it's not just a clever box that calls itself an intelligent computer it's going to be everywhere it be in the fabric of our buildings and everywhere that we we see touch and generally exist we currently have three four five six companies I would go through the wall who are making humanoid robots they all look pretty similar I'm sure they're actually internally radically different but for me the the big question is beyond the mechanical utility of these things which is something we'll come back to and talk about the complexity of it the thing that really interests me is the ghost in the machine the intelligence that is driving them hello and welcome to another episode of how to understand today your weekly insight into the technology that's shaping our world I'm Simon Windam Dave shton I'm Phil rhs today we're going to be talking about robots whether you see them as the Terminator in Waiting or humankind's best friend we're going to be discussing the latest developments and ask just how far things are going to go is Elon Musk going to create an army of Doge cyborgs to take over the world or are we going to go more in the direction of Apple's recent cute looking Luxo Junior style tabletop lab now anyone of a certain age will have been used to the rather crap looking robots that often got presented at trade shows as the future of housekeeping it was a dream that never materialized and even now I can't buy a human robot that will have a full CIP breakfast ready for me when I get out of bed on a morning is the promise still real and is it hyperbolic to be worried about heading towards the universe of iRobots or the Terminator Dave are we going to be seeing robots wearing an apron in our kitchen or serving or derves like we were promised in the 1950s or has our perception of how robotics can integrate into our lives been skewed far too much by science fiction well I think both of those are true actually um I I I think you can hold those two thoughts in your head at the same time the recent progress in humanoid humanoid robots is such that we may well be seeing them uh as as the primary occupant in the kitchen um but also a bit like AI itself it's not just an app it's not just a clever box it calls itself an intelligent computer it's going to be everywhere it'll be in the fabric uh of uh our buildings and uh everywhere that we we see touch and it generally exist so um and and the reason that robotics has suddenly come to life is of course because of the massive advances in AI that we've seen recently um and it's worth putting this in perspective because Robotics are to get a humanoid robot and let's talk about humanoid robots because they're the ones that everybody loves or is scared of um I think I think it's you know you were talking about the sort of crap robots that we we've seen where you know they wave their arms around a bit and they've got a red bleeping light on their head and and almost completely incapable of doing anything other than you know in a rather unseemly way tripping over the carpet and falling over um and yet and and you might remember was it 10 12 years ago 15 years ago the the Honda had a kind of humano humanoid robot called was it Asimo yes remember it and this thing just kind of TR along it kind kind of walked it even did a really kind of uh rather limited kind of dance that everybody thought was impressive at the time fast forward to today and you've got human or robots that can run jump do some assaults do choreography um and have like fingers with like 27 degrees of freedom so in other words they're they're getting close to the over overall functionality the overall mechanical functionality of uh a human being um so we we currently have three four five six companies I would go through them all who are making humanoid robots they all look pretty similar uh I'm sure they're actually internally radically different but the for me the the big question is beyond the mechanical utility of these things which is something we'll come back to and talk about the complexity of it the thing that really interests me is is the ghost in the machine the the intelligence that is is driving them and you you have to consider this in multiple ways because it's not just the ability for them to talk to us and understand when we give them commands it's the ability to control their own body and deal with their senses and I guess this is all down to um technology convergence the the speed of processes that are available now compared to what we used to have yeah absolutely right we would without the advances in AI we wouldn't have even come close uh to this although I have to say that Boston Dynamics have been doing some very good stuff for I think over a decade and and their original kind of big dog kind of robotic life forms uh were pretty amazing not least because they sounded like they had an internal combustion engine I mean they were unbelievably noisy and I assume that that is some kind of generator just because the batteries weren't cap you know it would take three steps and then stop but obviously batteries have been improving at the same time but what was impressive about those early Boston Dynamics um machines was that they could self-correct their stance and their posture so you you had these videos that always made me feel a bit uncomfortable in an anthropomorphically motivated way um when somebody would kind of kick one of these robotic dogs and it would stumble and somehow stay on its feet that kind of autonomic response is so important and that's hard enough with a quadraped when you get into the bipedal mode um you have to have incredibly fine control and and real time so when I guess the point I'm making is when you're building a robot you have so many layers and levels of automation that you handle even before you start talking to it or expecting it to interpret the world so I mean going back to the Boston Dynamics thing they they achieved an awful lot with probably very limited AI um and then they they they were they primarily powered I think hydraulically and that kind of reached the limit of its ability and and now they're back to kind of electromechanical um uh models at the same time as improving their you know brains if you like and the risk of going on too much here because you can't start on this subject and and and then stop without covering a whole load of other things but we I'll just mention them briefly and then we can come back in detail but um first of all we always thought it would be really hard to be able to talk to a robot and you can remember probably remember over Generations generations and at least a decade or two of people supposedly talking to robots and just getting a really well robotic response but not even a good robotic response uh it was almost like you do you remember the musical instrument called a melron which was the first ever sampling instrument and it didn't use didn't use um computer memory it used little strips of tape and when you press the key this tape which would have like one note of a piano or one note of a flute or a choir you press the key and it would start playing that piece of tape over a playback head um and it could do it kind of polyphonically because each know it had its own strip of tape well when you talk to these robots it's almost as if they're playing a strip of tape because their responses were so stilted and you can imagine a kind of primitive decision tree you know say to it how are you and it will say I'm very well thank you how are you today uh but you never got very far with it you certainly couldn't discuss abstract concepts with them or it goes back to anything like a a response that that a human would mistake for another human being whereas now we have our Transformer based llms these large language models that are extraordinarily good at conversing with us just over the last six or 10 12 months we've gone you well I guess current state of the art is probably the um uh the native voice mode in in the chat GPT Pro which can not only Converse almost with no delay but even put on accents if you want to so the reason it's called native voice mode is because it's not just a computer voice reading some texts it is actually generating uh the voice in real time to satisfy the the the the conditions that you've or the parameters you've given it so you might want you know might say to right read me the shipping forecast in a Northern Irish accent and it will do it um so given that that's one enormous problem of communicating with robots that's been solved yeah I mean on on that it is quite um good when chat GPC um does that although it does have a propensity to get stuck in an Indian accent um I I I yes the politically the politically correct Brigade will not be not be happy with with it I once asked to say something in Welsh and it it it did so initially um but then for some reason it went into a an Indian accent and wouldn't wouldn't leave it even when I told it to go back to an English one anyway Phil what's your what is your sort of opinion on all this I guess in my brain this breaks down into two things this breaks down into the AI part of it and it breaks down into the mechanical engineering part of it and I think the AI part of it is kind of fairly well covered and I think you know you can you could you can train an AI to walk a robot exoskeleton around that's no problem or at least it's you know a more solvable problem than it ever would have been so the mechanical engineering part of it is is the part of it that's that's interesting to me particularly because it starts hitting some quite fundamental problems I mean for a start the not problems but challenges the first part of it being what do you want this robot to do and when people say robot they think you know the Terminator it's a it's a humanoid but the only reason you would make a humanoid robot is if you wanted it to be able to do all the things a human can do in which case you're aiming to make a very general robot which is going to need an extremely General AI that that can deal with a huge range of things otherwise it's not useful if you want it to do anything more specific than that if you want it to Hoover the carpet you don't make something to push the hoover around you make the little spherical cylindrical thing that buzzers around on the floor because that's a robot and that does that job the only reason you want to make a a humanoid one is is for some reason you want it to be able to interface with all the things that humans interface with which is what you need if you want a general you know household help or whatever but is is not the case if you want anything more specific than that and the challenges of it and which comes back to what we were saying about the sort of internal combustion engine noises are that human beings and the Motions that we can make have a pretty they have a combination of speed precision and power that are actually quite difficult to build and one of the places that we see this is in some of the Disney theme park rides there's a Disney theme park ride which is based on the the Avatar film which has an avatar with it's a large humanoid thing and it moves like you would expect it to move but to do that they using the very Zenith of industrial robotic technology which has to be plugged into a very large power supply because to do you know this with your arm my arm's moving quite fast and my arm weighs a few pounds and you know the reason I say this is because I have at least some small background in building little props and things for for film and TV work sometimes increasingly historic now but I do and one of the the things that you frequently see in uh that sort of what you might call a light industrial environment is pneumatics because pneumatics can generate quite a lot of power we would build pneumatic devices that would have little CO2 cylinders in them to run them and You' run your CO2 cylinder out really pretty quickly if you have a CO2 cylinder and a a Pneumatic actuator that can actuate over say 50 mm 100 mm um and with a a few tens of pounds of force you only get 20 30 40 activations out of something like that before you've emptied your little CO2 cylinder and this is why these things sometimes have ended up with internal combustion engines on them because people go oh I'll use batteries the problem there is the same problem we have in electric cars over the last 20 30 years battery prices have come down and things have improved a lot absolute battery capacity has has improved but if you look at a graph of computer horsepower over that time it's an exponential curve because of more and then you look at memory capacity and it's an slightly flatter but an exponential curve and you look at storage capacity it's still an exponential curve you look at batteries and it's kind of FL it's not quite flat but it's it's not great which is why electric cars are still in terms of range compared to an internal combustion engine car not great and building some building what you want building Mr Data from Star Trek what does he run on that's like one of the big mechan techical engineering problems because he's very strong he's stronger than a human being he's very fast he can do all the things that we can do with the speed and the Precision more than us what does he run on what does he's also 400 years in the future but this is the thing you know and but you said something interesting though Phil which I think is really important and you're saying you know why build a humanoid robot to move a vacuum cleaner or trundle with a wheelbarrow um but what we're talking about is the you know in a really non-trivial way is the difference between a a specialist tool and a general purpose tool yeah so like um a calculator is is a niche tool a pocket calculator a computer is a general general purpose tool and uh I think I think you're sort of AI uh rber type robot is a a a specific kind of Niche thing whereas uh the entire retra of a a robot is that it's a general purpose tool and so yes it can do all of those but probably at 20 or 50 times the cost um but I I I just wanted to say so you can't go very far into a conversation about robots without talking something which sounds like horrendously technical but is is is like crucial and can be easily explained with AI and robots and that is that uh imagine that AI at the moment without hands feet head Eyes Ears senses touch and so on it's essentially a brain in a vat of oil you know it's it's functioning but it has no connection with the outside world and that severely limits not only its ability to learn but its ability to understand stuff at all about the world now the big the big sort of generative I mean in the sense of being open-ended and massive the big generative step that you get when you put your um disembodied AI brain and it's not a brain of course it's it's a computer but when you put it in inside a robot just imagine that the difference you've got there because you're not now restricted to training it in a a sort of uh data center way uh you you train it by letting it wander around and find out and sorry I was just finish this bit so by the time it can you know see something and like wonder what's that what's that orange light over there and what's that that thing oh it's a flower I've I've heard of those but I'm going to go and walk around it I'm going to touch it oh the petals have fallen off you you know and all of a sudden you have uh a robot was not just learning about stuff it's learning about it in the world its place in the world and that that is like a that that's analogist you you've heard of the there was a uh a lady from about I think probably 100 years ago called Helen Keller who was uh both uh blind and deaf and you know 100 years ago sorry I don't know the exact dates um that that was a recipe for a pretty miserable and limited life because you you you just couldn't learn you couldn't communicate you couldn't learn you you just exist on a level but you're idea of the world would be extremely limited because you couldn't go beyond your internal thoughts and your your Sensations so you you know you'd be limited for example to you know thinking hunger you wouldn't be thinking I am hungry you'd be just be thinking H hunger because you you haven't even got the ability to self-identify you you know you you and you don't need to talk about I if you if you're not aware of the world out there you don't have to distinguish yourself from the uh the outside world so but what happened was that people started to communicate with her by they taught her an alphabet then they talk taught her the meaning of the words by by just touching her hand uh and through that she said she suddenly had a realization that the world wasn't just internal to her there was a world out there and she existed in Contra position to the world and that sounds might not sound like a big deal but it is the biggest of deals it's realizing that it's not just you it's you and the world and other people and language and everything that so and when you put an AI into a fully functional robot um and it's it's called embodiment it it's it's been known about for some time now that uh to get a truly great functioning learning AI that understands the world it needs to be embodied it needs to have a body it needs to have eyes and ears and senses um and sorry I'm going to stop in a minute but if if that wasn't profound enough um we all have to go to school because we can't learn by someone else going to school for us robots will all be connected so they can all learn collectively so they'll learn millions of times faster than us yeah so so just sorry just Phil just want to just interject Phil I just to go back to your points on and it's kind of plays into a bit about how we think about robots um now pH you you mentioned about the the the Hydraulics powering the limbs which um brought up memories for me of the very famous hissing Sid demonstration on Tomorrow's World once which went very very wrong it's worth looking that up on uh on YouTube if you haven't uh haven't seen it but one of the things or aspects of research that um is going on at the moment is into um how to develop a a functioning artificial muscle a flexible muscle that can control Limbs and uh and therefore make a much more humanoid style robot now the materials to do this you know they're they're proving problematic to to do in any sort of scale they can do it on a microscopic scale which actually has side benefits for U medical robots ones that can go inside the body and they can change shape as they maneuver through the arteries and and unblock things and and that sort of thing and there's a lot of research going on on there they've done simulations they've had an AI Control this particular robot going going having to fit through different gaps and it changes shape um to adapt to the situation so you know we often think about the about robots as being humanoid or being like something from short circuit or the Terminator or whatever but actually a a massive amount of research and development is going on in terms of microscopic robots um that can really really help from a sort of medical um point of view everything from a you know moving light in a nightclub in the late 90s is a robot you know an industrial robot which comes back to what we're were saying about them being Tas specific something that Dave said made me think about something about these things all being interconnected I mean you know once here I am coming in with my jug of cold water for right now they're going to have to be because one of the most power hungry things you could put in a modern robot would be a Neal Network um yes there's lots of work going on with truncation and quantization and compression all kinds of things right now it's hundreds and hundreds of watts to run a a decent Ai and anything like a real-time manner as well it's really really difficult you you're not going to Chuck a GPU in the back of a robot because that's going to pull half a kilowatt and again what's it going to run on so is the brain of this thing going to be in it but that when we're talking about them learning I mean this is again become to thing about much more thing about um about AI than it is about robots in particular but it you know an AI can learn by looking at the world through things that aren't part of what we might identify as a robot it can look through cameras it can listen to microphones it can do all those things without necessarily having a body the question is do you want this thing to become a human being do you wanted to learn what it's like to be a human being philosophically or and how would you tell how would you be able to prove the fact that what does it feel like to be in that body I don't know what it feels like to be in another human being's body let alone some robot body so how do you even prove that some philosophical stuff that gets quite interesting as well but I mean that that sort of plays into how as well into how we respond to robots and I think this is why um the Apple Luxo junor lamp style thing that that that was unveiled this this week um is is quite an interesting one because it's one of the things that they're doing with it is they're experimenting with how this robot can exhibit character even though it's this just a lamp yeah just like the just like the Pixar Animation so it's doing stuff like if you ask it what the weather is like it will glance out of the window and yeah ex well it may be but the whole point is to kind of create an emotional connection with it even though it's just a lamp and you know how we relate to robots is is is going to be quite important I think in terms of like um in terms of you know making humanoid I think all of which is kind of circling the question of um Can a robot be conscious and what would it mean if it if it were well we're not we won't deal with that today because that's a whole separate conversation um and I think what what you're just saying assignment uh the sort of anthropomorphism is actually very important but all I was going to say is that so let's imagine we got you know the touring test you remember the touring test so the I think there are various definitions but the the core of it is that if you can have a conversation with a computer and not realize or not be able to distinguish it from a human being in a blind test then then you have to say that to some extent that is is int artificial intelligence well in many ways we're way past that now especially somebody who's not familiar with the latest developments in AI um and in fact Ray Kurt's file made the rather important point that in order to pass a a touring test these days you'd have to dumb it down yeah because you know you otherwise you'd be asking your um you know your your subject your experimental subject uh you know what's your favorite poem and it will it will read you the entire uh works of from from memory and probably do it in whatever accent you ask it to do it in um but I I think here a philos opical thought experiment um which is if a robot uh it does it doesn't have to look like a human being but if a robot could behave exactly the same as a conscious being would but you know in every way how would you you know how could you tell it was not conscious or would you would you know it was conscious or not is that is that sufficient cause to say that it is conscious if it behaves exactly as if you would expect a conscious entity to behave well well that's going to cause some um very interesting uh moral dilemas if anything gets to that stage if if we're talking about robots being sent into battle um you know um the military is obviously very interested in in robotics and as you alluded to earlier Dave this idea of um connected learning so you know everything's feeding into everything else you know it might sound hyperbolic but is it really to to question the idea of effectively a real life Skynet stroke Terminator style scenario coming into into into being completely reasonable I mean and this is the thing none of this requires robots either I mean a lot of these questions about AI in general um whether or not it's but apply it to robots um assuming this thing is fully autonomous and self-sufficient I mean it is actually addressed in The Terminator movies what do these things run on Well Fusion Power of course because it was the '90s but assuming you can't run them out of batteries then I mean I think one of the most terrifying things I actually discussed this with a colleague of mine about a science fiction script we were working on that what's the scariest thing you can imagine being a robot okay Big Shiny steel skeletons they've been done um and then we thought we looked at this bit of footage of an AI flying racing drones and we thought well if an AI decided to run that drone into you it would and you would not stop it assuming that you couldn't dive into a vehicle or something if that we've seen this in um Ukraine recently when you on these things they are horrifyingly effective an AI could run that thing into you if you were caught in the open it would get you and there would be nothing you could do about it you know if it wanted to avoid being shot at it could and it would be very hard to avoid AI safety again this is an AI thing not a robotics thing but AI safety is a real field and a real field of research and that's normal that's absolutely something we're aware of I hope we're aware of it enough and I do wonder that sometimes the companies who need want to make a living which is completely reasonable will prioritize making a living over making a living safely which is not an unprecedented concern and is not something we haven't had problems with in the past um if you make it fully autonomous then it's very hard to shut down potentially if it's a racing drone the racing drone doesn't have the AI on it so you can go to wherever the AI is and unplug it and that's you know there there solutions to that um but in extremist if you can get these things to be you know self-sufficient which is being talked about in terms of autonomous aircraft certainly you know that's a robot um it's not a humanoid robot it's an aircraft shaped robot looks like a fighter jet in fact the ones that I think being Direct developed here in the UK a project called tyrannis is designed to be run either crude or uncrewed so it has a cockpit and a seat and and if you're a military project then you really don't want the thing to be messed with externally because people will try to um so what are are you running the the system on board are you running the AI on board this thing in a j fighter few hundred Watts isn't no problem you got that much power available to you is that okay this thing is autonomous it can go wrong a Terminator scenario is entirely possible I mean yes and if to certainly to the extent that if people are not taking that seriously then that's worrying and I think if you look at the rate of change just over the last year especially with robotics it it is accelerating it is inevitably accelerating and I think so it's very hard to it's very hard to extrapolate from where we are now to where we will be even in 6 months never mind 6 years but um I I think you know we brought up the question will it have the AI on board I think it'll have some AI on board uh but it will mostly have uh recourse to a AI in the cloud as indeed we we will when when our our brains start to uh essentially treat cloud-based AI as a brain prosthetic so you know one of the most interesting areas of research uh is is the the brain computer interface and and again that's seeing progress and just like with the early MRI scans uh they're pretty low resolution but the resolution is increasing all the time and again that's an exponential thing you only have to double the linear resolution to have four times the amount of data and probably another order on top of that in terms of its usefulness and interpretability so we we're very much in a virtuous circle at the moment with all of these Technologies but I think robots will be transformative because they'll be out there in the world I just want to say you know not not to put a damper on the robot thing but just to Ram home how far they've come um people that have never met me probably wouldn't know that I I actually have two prosthetic legs I was born with both legs missing above the knee so um I'm very lucky to have two so-called robotic knees um are they really robotic arguably but arguably not uh depends on context uh do they Propel me forward no they don't what they do is they control What's called the swing phase which is when you take a step and uh without any control your lower leg would act a bit like a pendulum it would have a fixed rate of Swing which is great if that's the speed you want to walk at but it's no good if you want to go faster or slower so these things uh electronically control fluid based or in other words hydraulic dampers to control the ballistics of your of the step you take um and it's it's pretty effective but it's it's a tin tiny tiny part of an entire human humanoid robot it doesn't even Propel you forward it just has a very small battery that controls a few valves that that can use this fluid to damp uh your movements and to overall make it more efficient and it's fantastic and uh uh actually very expensive luckily for me provided by our National Health Service um and of course I got two of them but um so to to me that that just illustrates how incredibly Advanced the current generation of humanoid robots are and in fact you're probably used to seeing I think I think uh the Tesla op Optimus robot still walks like it has some kind of chronic constipation you know it's I don't understand why it's so bad given the other other um robotics developers have made much better ones I mean I see it this year uh the Beijing uh half marathon is opening up entries alongside humans to the various robotics companies around the world what's really interesting about that these robots that walk with a much more natural gate as in git um they their legs are much more akin to human legs in that they when they they'll take a step and the lower part will swing forward and then it will lock and then they put their weight on it and it will be straight none of this kind of bowlegged kind of action the the reason they've adopted the kind of bowlegged stance is because it's easier and quicker to make unexpected movements but it's very inefficient for walking whereas the the the latest uh Japanese the unitary robots you've seen them running and walking with a very natural gate and they they have the kind of locking knee but still managed to be incredibly responsive so it's like almost every week we're seeing a major breakthrough in the mobility of robots I mean I think you know with with the ultimate conclusion as I um was talking about earlier in terms of organic artificial muscles and you mentioned the interfacing of of um sort of electronics with with uh human with the human brain and and and stuff like that well you know if we take take this to its kind of logical conclusion we are really talking about something akin to Bishop in in the aliens films really an artificial human life form and then obviously the moral quandries uh become very much like like the movies um and you know in terms of space exploration what would you rather send out into space to go and explore you would send something like Bishop because he doesn't age he can go as far as you want and but the thing is you talk about something like that and you go well why would you send that when we sent robots into space to explore already why would you send something humanoid well the only reason you would do that is because you have an expectation of its ability to explore places being based on the ability of a human to explore places and you know it's frustrating that these these Ro these Rovers they have on Mars are frustrating they're slow they have to be programmed every every every wheel turn they take has to be planned hours in advance and sent up and carefully executed partly that's because of the time delay in transmission reception but you know um the reason you would send Bishop out is because you wanted something humanoid because you wanted to have a human style ability to explore but that might not be the best way or to interact with other humans well possibly but then you know that does it become weird you know there are all these questions like should should you do that and how would that be and people make movies about that and that's you know that's where you get X Macker and so forth but yeah it's interesting yeah well I mean just drawing all these threads together um uh and really the main thread is is again exponential progress so I I think if we look back at this uh uh episode in six months it'll be like well you know they what what they're talking about is prehistoric uh and and uh the it's the networkability of these things even though they're mechanical their ability to share information about what they've discovered about the world and uh that's in parallel with the fact that a lot of the work Nvidia is doing at the moment and other other companies uh is building Virtual Worlds within which to train Ro robots and because they're virtual they can run millions of those in parallel you I mean this is this is clearly um a topic that we're going to be returning to um at some point because there is a lot um to discuss and there are probably some very specific areas that we can go into as well so I think we'll we'll leave it there for now um Phil you need to get off quite clearly um so yeah we we'll leave it there and we will uh reconvene uh for uh next week's episode so thank you very much guys um and we'll speak next week yep see you bye [Music]

2025-02-20 18:43

Other news