Episode 109
The Anatomy of a Robot
About this episode
In this second installment of our three-part robotics series, we dissect the inner workings of robots and explore the components that bring them to life.
While science fiction often romanticizes humanoid machines, this episode is firmly grounded in the real-world anatomy of robotics—how artificial intelligence (AI), mechanical parts, and software work together to build functional and increasingly autonomous machines.
From sensors to actuators, we break down the building blocks of modern automation.
Key highlights
- Brains, Controllers, and AI Integration. At the core of every robot is its “brain”—the controller that processes inputs from sensors and determines actions. We examine how hardware limitations have historically held robots back and how recent advances in miniaturized processors and edge computing are changing that. The team also compares the computational capacity of robots versus humans, highlighting how AI enables faster reaction times and specialized performance in specific, task-oriented domains.
- Sensory Systems and Real-World Awareness. A robot’s ability to navigate its environment hinges on a blend of advanced sensors, including cameras, LiDAR, infrared, ultrasonic, gyroscopes, and more. These devices serve as robotic "eyes," "ears," and "skin," enabling high-precision perception. The episode dives into how robots not only perceive physical reality but also extend beyond human capability through internet-based data input—offering a new kind of digital "sense."
- Muscles, Actuators, and Energy Constraints. The “muscles” of a robot—its actuators—convert processed commands into movement. While humans outperform robots in versatility, adaptability, and energy efficiency, machines shine in strength, speed, and precision. However, major challenges remain, particularly in energy usage and power systems, with robots still heavily reliant on limited battery life. The discussion compares robotic movement and endurance with biological efficiency, revealing both strengths and limits in today’s designs.
Robots may seem like futuristic sci-fi marvels, but understanding their anatomy—from AI-driven minds to real-world sensors and actuators—makes them less mysterious and more tangible.
While still far from matching the adaptability of humans, robots are becoming increasingly powerful thanks to rapid progress in robotics, AI, and mechanical engineering. As we build toward a more automated future, knowing how these systems work is key.
Join us next time as we explore how these technologies are being used in the real world—and where they might take us in the decades ahead.
This episode is sponsored by the VeChain foundation. Learn more about VeBetterDAO here: https://vebetterdao.org/
Transcript
Auto-generated from the episode audio.
So, welcome back to the Entangling Web 3 podcast. Now, this is episode two of our three-part series, special deep dive into robotics. So, in the last episode, if you haven't listened to it yet, I would highly recommend going back there because a lot of what we're going to talk about today probably won't make any sense. Um, but yeah, last episode we looked at how robotics has evolved from a historical perspective. We looked at some of the earliest kind of theories and conceptual ideation around robotics and we went through all the way from the kind of practical application of robotics and how it's emerged.
Everything from the works of people like Isaac Asimov and how it's broken through into cultural domains as well as the actual practical applications and breakthroughs. Everything from Henry Ford to the proliferation of factory arms, Rooma and Asimo. So yeah, we ended with kind of robots stepping into dynamic environments, things like drones, quadripeds, and beyond. Uh today's episode now, we're going to start to define exactly what we mean by robotics. So what are the core components of robots, what makes them work, and and what they're actually being used for right now from hospitals to farms to factories and homes.
So yeah, let's get into this one. You know, there's so many different types of robots. maybe it'd be useful to kind of define the core elements of all robotics and like what is kind of happening across those core elements. Um, and the first thing which I think is obviously probably one the most important from like the human perspective is the controller which kind of can think of as like the brain and this is like the core processor of the robot. It receives all the inputs from the sensors, processes the data and decides what action the robot should take.
And this is kind of like you know the one of the booming signs that we talk see um what's his name Jason Hung always talking about how you know they have semi they have processes that can process more and more and more on a smaller smaller scale like the idea of doing mass amounts of computation in a very small and lightweight chip real time is really essential for operating basically operating these brains for robots actually go out and be useful. Yeah. And this is I guess you know when we talk about all the scaling laws in in technology, this is one of those areas where as you said you you can you're being able to simulate more and more of that computation capacity or not simulate again but just you know achieve these levels of computation in smaller and smaller devices which lend itself much more to robotics right so like I think the iPhone is a nice example like the the the the increase in the number of flowing point operations you could do
on an iPhone right the the the the chips on on those themselves. Every year the iPhone is basically the same, but there's normally like a five to tfold increase in what you can do yearon year, right? So, we're at a stage now, you know, we're talking about the early 2000s when people are starting to look at humanoid robot robots. Now, the capacity to do that and jam everything into a much smaller package is much more feasible. Right. Yeah. Exactly. And it kind of like these the hardware setups which is different to the software defines like how much computation you can do which is like how much sensory input you can take and still kind of handle it and do something useful with it and how long that would take to do.
So I think like one of the interesting things is this is these are determining how fast and intelligently a robot kind of act and respond. And one of the one of the insights or one of the trends we're seeing more and more is that a lot of these controllers kind of operate as like a in a distributed manner where some will be done locally like you can kind of think of that as maybe like your um kind of subconscious reaction where it happens like immediately whereas like you know something's coming at me a car I need to jump out the way and that would always be done locally but some can be handled consciously and maybe offloaded to I don't know the cloud or something elsewhere like a network and that would be like you I think one of the things we saw the Tesla robots, the Optimus robots, um they need to be kicked to Wi-Fi and you could plug into your chat GPT and still create, you know, normal human responsive.
You don't necessarily want that to be done on the robot itself. Um and like just to put this into comparison, I think it's always good to compare to like the human brain to make it more like kind of relative and understandable to us. So the human brain processes roughly one exoflop, whatever that means. Like that's a it's a big number, right? Um across 86 billion huge 86 billion neurons. Um the most advanced robot controllers right now operate in the hundreds of teraflops range which is a million times slower than the human brain.
However, like a kind of like the human brain is, you know, there's this kind of thing where it's like it's quite inefficient. Like we're not as task focused where a machine can kind of excel in very specific high-speed narrow focused tasks like object detection or motor precision, all this kind of stuff. So, it's got all that compute power focus in one narrow field, but it is still like 1 million times less computationally capable than a human is right now. Yeah. And this is interesting, right? because um I guess you know I if we're going to cover the full picture of of robotics you know we're going to get on to the other two pieces but we're effectively talking about the sensors then the actuators on either side of this brain right so the whole point is the brain is is or the the the the control plane basically of these robots is how fast can you go from input to output basically right and I think there's there's probably a distinction to be made on
the brain part here because we also talk a lot about AI, right? And that is about prediction, about trying to generate something new based on inputs and outputs. And if you think of a humanoid robot, I I think of, you know, my smooth brain thinks of it this way. It's like uh it would not be too hard to simulate me, but I'm thinking you could chuck chat GBT or an LLM into the head, right? And you could simulate speech quite well now, right? you we have all the models, all the tools to simulate the kind of the the outward appearance and perception of it being human.
But the difficult part now is how do you simulate all the the rest of it, the tactile stuff, the movement, the decision making on physical on the physical domain, I guess, is what I'm trying to say, right? Because and is that am I right in thinking that's still the bit that everyone's trying to work on? They're not working so much on the kind of, you know, that's been pushed to the side of AI. Robotics is is essentially the AI of the rest of it, the moving and stuff.
Is that is that a fair assessment? I think so. I think that like software development is kind of handling not handling that in isolation is kind of like it can do that without the need for you know the humanoid style sensing stuff. I think it's actually the sensing and output the operating in uncontrolled environments in the real world that is really difficult. language stuff as we've kind of seen from ChapT is quite covered and like typically not super important like it's not life or death type things where it's like speech in speech out style but the idea of like obstacle avoidance the idea of being able to you know navigate through maps go up a staircase that stuff is the really tricky stuff and the kind of yeah definitely the harder stuff because hardware right it's hardware simulation versus hardware and that stuff is very tricky but like we're not talking about software yet we're purely talking about the hardware aspect of these chips right now and like it's crazy to think of the fact
that humans right now are still have a million times more computational. It's a a really silly way of putting it, but it is the way I want to put it. So, but the processing power of these hardware chips is catching up very quickly. And I think another important aspect is a human brain is individual like once you die or something happens to you, it's gone, right? But the idea of like shared and stored learning across you know hardware chips is very powerful to have like you know I can copy paste the learnings and the brain of one robot to every million robot and have a network that kind of controls distributes allows them to work in kind of like a hive mind together and that is not possible in the same way with humans which I think is a massive value out of doing this on robotic systems.
Another really interesting thing is like the speed of thought. So the speed of thought or the say speed of thought the speed of like you know signapse compared with like electric circuit is around a million times slower. So the reactions of a robot from like visual input to reaction versus a human is around 40 to 50 times faster. And I think the idea that's like the speed of possible computation not the amount of computation but the speed that computation can happen at is also far favorable towards Exactly.
It's far favorable towards robots. Again, another big win for robotics. Yeah, that's interesting. So, it's like there's actually the at the the in biological terms, we have certain limitations and the computational power is, you know, biology wins right now. But in terms of latency, that's actually that's a win for machinery over robotics, which makes sense. I think if you want a concrete example of this, you should go and watch the the robot goalkeeper. Have you ever seen this? Oh, yeah. I've seen it. Yeah. This this is kind of a way of visualizing that, right?
Because the robot goalkeeper they you know you get Messi and Ronaldo try and take a penalty against it and it'll save it every time like nine 99 times out 100 because of that I guess that that latency right because it can just see the input and it's like well I'm going to go right there and save it but the human brain doesn't quite have that again like and when you kind of want these things in life or death situations like where like on a construction site even like as a guide to make sure that someone doesn't get hurt like that kind of reaction speed is very valuable.
an autonomous car, right? Exactly. That kind of thing like it's the most the hardest thing for autonomous vehicles is humans. The unpredictability of humans, but even then like their reaction speeds are so much faster than what a human can possibly do. The question is like who's liable when something goes wrong and there are some scenarios where there's no good there's no good case, right? There's there only the really bad case and then the slightly less bad case and then what happens then? How does he make the decision?
Again, the laws of robotics kind of come into that hopefully. Um, so the next big thing is sensors. You talked about that's like the information that is coming in to the robot and how it interprets the world, how it understands the world. And there are so many different things like humans with the five senses. Um, but you know that is actually a small part of what is really going on in the world. Like there are so many more things that can be detected and robots do have the capabilities to detect those.
You have vision which is obviously so important for humans like the idea of cameras. They can use cameras that obviously a greater spec than what we can see. But you also have things like liar which is like laser sensors which are very good for like exact estimates of distance to things which is like what you can use in you know autonomous vehicles very accurately. Like lidars can work in fog or you know poorly lit conditions like dark nights and that kind of thing. You have 3D cameras that can do similar things as well.
And then the more kind of like the things that we kind of think of as less robotic in a lot of ways. like touch and pressure sensors which you can use for like graspers and that kind of feedback. Pro more proximity sensors like infrared, ultrasonic, you know, batches, that kind of thing. And also like inertial sensors that we have in our phones where you kind of have like the gyroscope that knows how much is being tilted. These are so important for movement and knowing like my positioning relative to what's going on the floor or something like that.
Yeah. You look at something like a gimbal, right? How good they are. Now like there's I think again the iPhone is a great example. If you just look at the number of sensors on an iPhone in 2025, it's it's it's like 10 hundfold what it was when the iPhone first came out. Like it's insane the the the granularity of inputs and the different range and type of inputs you can get now. Um I don't know if you class this under sensors. Um but I would also maybe put something like the internet here as well, right?
So robots have the ability to take inputs and stimulus from external sources, right? Like you can connect it to an API. You can connect it to the internet to get information and things like that which again the human doesn't have. So maybe that's another a lot a loss for biology in this case. Maybe that's really interesting. I never really thought about that. That is quite cool. I think like when we get our Elon Mus style brain chips will have that like sensory input but like I can't even imagine what that means right now because obviously our all input from the world goes to our brain via one of these senses.
But the idea of like automatic upload is bonkers that it can just skip these senses and maybe that is a sense in itself that we don't have right now. Um I think most robots right now tend to use or like I say most robots most of like the highgrade like out there robots tend to use a blend of these like in the same way that humans do like your blend of these sensors. Um and then they like have a semantic perception that allows a robots not just to see an object but understand what it is and what it should be handled for.
Like you imagine like a labeling system. It's like I see car, I know car, go fast and hit me, avoid, you know, that kind of thing. It's that level. And someone has to go in there and label it. And obviously AI models are great for that kind of stuff. Handling vast amounts of data and, you know, creating rough labels of what should and shouldn't be done, maybe even synthesizing labels. Yeah. So that covers nicely, I guess, the input, what goes in the sensors. We've covered the brain.
How does it get processed? How fast does it get processed? Then the third part is going to be the actuators or I know you like to call them like the muscles of the robot, right? So what what it can actually do and I'm going to predict here that uh that biology has another another disadvantage compared to machines because again as we said that's the whole point originally of lots of the you know primitive robots or machines is to is to do things that we physically couldn't do because of other biological limitations.
So yeah I guess these are going to be pretty good right? Well it's actually interesting that it's maybe doesn't necessarily hold. So like the human arm can generate up to like 15 Newton meters of torque. And when you look at like some of the the major robotics arms, they go well over that like hundreds of newtons, kilon sometimes, whatever. But they're far less compact. So like the muscle to power ratio, power to weight ratio of like human muscles actually outpace a lot of actuators. It's just on robotic scale, you can just go much bigger, right, and scale it out, which isn't possible with a human unless you're Jack Davis.
Um but like it's the the real interesting thing is like the joint adaptability of humans that it can do so many things in so many different ways whereas robotics it's a lot more defined but then robots can go much more precise specific controlled humans are more analog whereas robots are more digital. So there's kind of like a I wouldn't say one or the other is necessarily good or bad. And again, we're taking so much inspiration from biology now inspiring robotics and actuators to work in a biological manner and that is definitely helping.
But there's a lot of other like big pros and cons that go to biology. The idea of like adaptiveness like I can build muscle if I exercise or lose muscle and that's quite good in like saying conditions 10,000 years ago I don't know how much you know I should consume or what my energy you know dissipation should be. I can self-heal. I can self-improve. My mobility is better. I can bend in many more ways than a robot really can right now. Far more ways than a robot really can.
And again, like I say, there's lots of inspiration that roboticists are taking from nature. And that's quite important because nature has really fine-tuned a lot of these things and has many different specialties from like how birds fly to how ants move and how ants have such a high power to weight ratio. How do they do that? And then humans obviously are one of the most efficient animals in the world for like energy dissipation and like resting energy dissipation. Yeah, that's really interesting actually. I hadn't thought about it in terms of you know some of the pro the pros we have.
I was assuming that this is going to be a hands down loss for us humans. But yeah, I guess the self-healing thing makes sense and maybe what I'm gathering from you is that it's almost like the the general purposeness of a human body or you know a biological body that we have in nature. they they're so general purpose and so adaptive to environment that that is that's an incredibly difficult thing to simulate because when you build a robot you must I I presume you're tending to you know you're limited by the materials you put into it at a given time right and what's available and you're building it to be able to solve X problem you know maybe we're trying to get that more gen generic I think but it's still going to be limited by how does it react to you know unfamiliar circumstances and then you know if if you haven't built have you haven't put What's the phrase?
You can't take the eggs out of a baked cake. Right. It's like once once you've built the robot, it's built and it's going to be difficult. You need to upgrade it. If it comes across a challenge, it can't. Whereas human bodies, yeah, they can be adapted. They can uh, you know, build themselves up, heal themselves when they need to. But then humans have a, you know, a shelf life and it's even harder to maybe change things in a robot than it is, sorry, even harder to change things in a human than it is a robot.
And on that like adaptability, we talked about efficiency. a big thing in robotics power systems. These things are typically very, you know, energy intensive to run. Like lithium ion batteries have been incredible, like really dense amounts of power in kind of quite light um containers. And obviously, you know, these store fuel in in the entire system and that fuel is delivered to motors, sensors, process absolutely everything to run it. But again, like when we look at comparing humans to robots right now, a human at rest uses 100 watts of energy and that's less than a light bulb, which is incredible to think of that.
Whereas robots are typically like five to 10 times more intensive than that and lasts one to two hours, three hours tops per charge. And if you thought like they had their entire battery pack on their back so they could last 24/7 or whatever, then they would be far less effective. So again, humans are very good at well very good at energy conservation and very efficient at how they use energy. We recharge with food and sleep can be done on the fly whereas a robot needs to be plugged into something and kind of has downtime in between.
Yeah. And that's always been one of the things I've kind of thought would be the longest lasting limitations of robots is, you know, how how do you keep them up? How do you keep them working? Um, battery technology seems to have been, again, from a layman looking, I'm not an expert, but it seems like that's one of the absolutely the hardest things to improve. Again, using my iPhone analogy, the battery doesn't improve that much uh as as much as the rest of it does over time, right?
And it's still it's still one of the big limitations. It's it's one of the issues people still have with electric cars, for example. I think today's episode is sponsored by the Vchain Foundation. Now, we often talk about how emerging technologies like blockchain have the potential to revolutionize sustainability and Vchain is at the forefront of this green movement with initiatives like Vet Dao. Let's take a closer look. Imagine tracking small actions like recycling at low cost making them valuable assets. These verified records build a massive collective impact benefiting people through their own actions.
Sound a bit complex, but Vchain made it fun. Explore our apps at vb better.com. Vchain's goal is to power a new era of sustainable solutions. So if you've got an app idea or just want to earn rewards for small sustainable actions every day, go check out vetero.com to find out more. And now back to the show. Okay, so the fifth core component of robotic systems is the software, the mind. Everything we've spoken about so far, the first four components have been quite physical, quite easy to understand, quite easy to relate to.
But this is where it kind of gets a bit abstract and it's probably why I left it last cuz I wanted to kind of spend a bit more time in it because it's also one of the most interesting and probably fast developing areas right now especially with the boom of all we're seeing in terms of you know Gen AI and LLMs and all this kind of stuff. So the mind of a machine like it's so abstract how do you actually kind of think of it? We think of it as software right and this software is what really breathes intelligence into robot.
It turns all the, you know, the hardware that we've spoken about in the previous four components into behavior connecting, you know, the sensors to the controllers and the actuator. It helps the robot effectively make sense of the world and then do something like act accordingly. And you know, if the we kind of described you people might be asking what's the difference between the mind, the software, and the brain and the hardware. This is my question. Yeah. Well, it's like the controller is the physical processor. It's kind of like a CPU or an AI chip.
It's like what's the potential? It's like the brain itself is probably like the equivalent of the hardware, but the software is like the logic that runs on top of it, right? It's like everything from the low-level control loops to the high level AI plan. It's kind of like the neurons and how they connect in a human brain. I kind of get a little a bit vague between those two things, but you can update software very easily, but you can't update hardware as easily. Okay. Yeah, that makes sense because I had this same issue when I was looking.
And I was like, how am I going to describe the difference between the software, the mind, uh, and then, you know, the the brain we describe as the controller? I guess maybe maybe I was thinking of it more as like the brain part is almost like the nervous system. It's the bit that's organizing information, sending out signals, kind of causing stuff to happen. And then the mind is you know that that gray matter that stuff that is harder to understand what is actually going on on like a mechanical level but we understand it's like where all the the the organization is happening right and it's taking it's the bit this process it's it's processing things and and learning how to do different tasks right but but again the mind and the brain they're they're a difficult distinction to make in humans right that's a big thing I think it's because we're trying to humanize these things that it gets especially complicated it kind of makes us self-reflect a little bit.
Whereas like if you were to say you know the chip my laptop in itself if it had no software running I could quite easily distinguish you know what is a chip in that what is the hardware component and then I upload the software the operating system and again I can quite easily distinguish between different types of software so I know there's a distinction there it's because we're trying to humanize these properties which makes sense as well that we can't really apply it to our own mind like we tend to think of the brain and the the the mind or the way we think as the same thing but I think more and more we are starting to disciplines like kind of separate those two concepts and the way you think obviously is affected by the structure of your brain but actually the things that are flowing through your brain are kind of independent of that and the way in which your brain works is independent of that.
Um yeah, it's kind of like the nature versus nurture there. It's a similar type of paradox, right? It's quite tricky to get your head around, but I think in terms of robotic, we do tend to separate those ideas. Yeah. I And also, I guess you could say the brain is something that we only use a certain percentage of the potential of, right? Like the the physical hardware in our brain, the wet the wetware has a certain capacity for computation, but we don't necessarily take advantage of it all.
And maybe that's part of the software. as part of what our learning like is how the neurons fit together I guess right so be one of those people that's like we only ever use 10% of our brain and if you use 100% you're like levitating objects and all this kind of stuff I feel like sometimes we're using significantly less than 10% on this Alec but yeah so so tell me what what does it look like then so when we're talking about the software the kind of the logic that's going on inside these anthropomorphized robots what kind of software are we talking is this going to be like just static pre-programmed logic that some human beings in a lab have decided this is how the this robot should operate or is there something more kind of technical going on there?
Well, that it's like again like I'm comparing it unlike like the physical components like the motors and the sensors that we've talked about software is the most flexible and kind of fast evolving layer. It's the thing that in you know in the recent in recent times has advanced probably the fastest especially with like say the LLMs and foundational models that we've we've kind of seen coming out but like there are some quite defined aspects of software and I guess this kind of works in humans as well like we're seeing more and more parallels with how a human brain or a mamillian brain or any organic brain works is you have certain aspects of the software like the low-level control stuff and that is like you know say you know you might describe that as like the subconscious the things in a human brain that I don't know I'm controlling but they just happen and they have to happen otherwise I die for example my heartbeat all this kind of stuff things you don't have to
think about so the the low-level control stuff that probably won't change too much in these systems because it's dependent on the hardware unless the hardware changes there's no need for that so that's like coordinate joint position for example you know when I know my arm is at 90 degrees versus 85 degrees those kind of low-level control systems the feedback involved in that for balance and actuator signals that's Like that's the one aspect that isn't going to change. It's quite simplistic and is related very strongly to the hardware itself.
Interesting. So would you say for this um this kind of software component? There's there's a combination of I'm just trying to kind of synthesize what you're saying, but is there a combination of almost like static, deterministic, simplistic things that can happen like you said there and then on top of that you have the the more kind of stat um probabilistic AI type things here, right? Because that's all based on probabilities. And when we're talking about AI in the context of robots, are we also talking about kind of more classical I don't want to say classical AI because it's you know it's still pretty new but like neural networks is what AI used to be right before Gen AI and to to a degree it still is but it's kind of wrapped up in other stuff but neural networks have obviously been based on the human brain right so it's part of this evolution of anthropomorphizing things which is very much linked to robotics so yeah how much of the the AI stuff is going into
this software and like the classical AI Oh yeah, that's the thing is like so you desri and we could kind of get into the realm of like nothing is really deterministic right there's a probability that's not it's not practical to go down that route but like there's the lowlevel stuff which is you know quite easy to understand and then as you kind of go up the technology stack and you're going into okay the sensors how am I processing say images how am I labeling images how am I clustering images that kind of stuff is like you say quite traditional AI the idea of like determining patterns and creating you know probabilistic I guess label for that based on say pre-trained data and whatnot.
The other thing is like we could think of all the sensors I'm getting all this object information like a 3D scan of the room. How do I relate that to useful things? It's not very useful to just say there's lines everywhere but to say oh those lines look like a door based on my previous knowledge. Again that's a degree of there's AI doing that. But like the kind of the one of the most interesting things is probably the classical things is taking that a step further.
It's the idea of planning. Like it's one thing to get the sensors, get the information, turn that into a basic understanding of the room and again there's AI, traditional AI involved in all of that. The next step is okay, what do I actually want to do in that room based on like my objective programming? It could be like you know you are a firefinding robot and you've got to go search a map and find the fire. So like I know where I am, I know what's around me.
The next step is okay, what do I do in that room? And there's things like slam that we could go down like self-loizing systems for like what I want to do next and mapping systems for what I want to do next. But that is like a very traditional and very well understood like 20 30 40 year old AI problems. How do I go? How do I actually achieve my objective based on where I am and what I know of my environment like I should go in the door for example or I should follow the heat map.
And they're like as these systems get more and more complicated like they have so many different say sensors feeding into them the the the maps are becoming more and more dynamic. There might be multi- aent systems more and more you need say very very heavy computation to actually solve these problems properly and in the most effective manner. Got it. Yeah, I think that makes a lot more sense now. Thank you. So it's like when we're talking about AI in this context, it's still very much predictive, right?
It's like predicting what is the right action to take. So I guess this is also where the the task orientation of the robot is coming in, right? So it's been given a goal of some kind and the software is using this predictive system to work out what you know in the room case the best way to to clean the floor, map the floor out, exactly that kind of stuff. Um and I guess this is quite different to again we talk about Gen AI all the time.
That's like it's a predictive system but it's it's kind of natural language processing, right? It's like it's limited to processing a certain set of media types. Whereas this kind of AI and the predictive models that they're using. These have been built to work around all sorts of different inputs as you said, right? Like image processing has been around for ages and ages now. So it's quite a different side of AI that that we haven't experienced so much, right? In in AI age. Yeah, exactly. I think that that is like it actually is far more intuitive than Gen AI.
Like there's very few people that could really understand because of the size of the data included in every single prompt and response like how exactly it is coming to a conclusion when I send a prompt how exactly is it actually determined that that that is the answer. But these systems are far more intuitive especially from a practical sense like you have a basic model of you know okay I know what I am and I know if I need to go forward then I apply a little bit of motor force to my two wheels and I will go forward and I know roughly if I apply a little bit of motor force I might go a little bit of distance based on you know the type of floor I'm on the type of slip there is all this kind of stuff and then as I'm like I will do that input get an expected output and I will measure that expected output based on my understanding of the world the map around and I'll localize again
and say was that about right and if it is I'll do a little bit more if it's not I maybe apply a bit more force it's kind of very intuitive how these systems work but when you talk about like geni and responsive and LLMs like that stuff is nonsensical to a lot of people even even us included I think Jack but that is such an added it's an incredibly intuitive not incredibly intuitive sorry say that back it's incredibly useful add-on to a lot of these systems to get that you know input output and a human readable response whereas Before it would been like I have to give a GPS coordinate to go here, here, here and it's very technical.
I have to code that in, hardcode it in potentially. But now to be able to give that in human readable format and it can digest that into something useful is is it's a bit of a game changer. Yeah, that's interesting. You're also making me think now that like AI in the robotics domain is it sounds like it's using quite a narrow context window. Is that fair to say? like it's using its local surroundings to make decisions and is and if you think of like you know I'm sure there's going to be a degree to which these models have been trained on external data and test and trials run previously but it's still within a narrow context of the same kind of events happening.
So that's what you know it's doing as a highly specialized job in many cases based on it's highly specialized training and this localized context whereas Gen AI is like taking the whole internet and then you're asking it a question right which is completely generic and in the background. I think that's that is a really interesting point. I think it's definitely a more specialist um data set like it's a much more narrow data set but then it's also like if you're thinking about how this is applied to say a drone system that is spinning like thousands of RPM and you've got many different you know many different angles and like it's kind of it's reacting to milliseconds sometimes like it has to happen it's not that it's like there's obviously a much smaller data set but when you're computing that to the millisecond and you're having like thousands of RPMs and thousands of different parameters like it can be very complicated as well.
So it's like the data set is much more defined and narrow but it's happening at much faster rate. So the context window is probably similarish style maybe but a lot more specialized. Yeah. Interesting. So like the amount of data but it's kind of Yeah. It's like the the narrowness is there but the volume and the complexity and the intensity of the computation is still incredibly high. Yeah. Exactly. That's a good that's a good way of describing it. The genai obviously the boom has helped with a lot of reasoning and language which is good for input output and some basic stuff like you saw the Tesla Optimus robots like one of their demos how true it was or how actually like not fake it was on site was that people were asking it real-time questions it was connecting to Wi-Fi and gave them like an actually decent kind of chat GPT style response which is quite useful but in terms of like the motor systems the control systems the real world planning stuff that's still really
hard like jai isn't at the stage where it's solving in these problems like it probably will it will never be like those classic Gen AI models that just understand language that can solve these. Sure they can help with the coding but you still need like copious amounts of training like they can set up the model say this is how you train this system like this but then usually testing testing and you can do a lot of testing in simulation but you know you then you have to go to the real world and be like how do these motors actually work?
How do these systems actually work? What's happening when I take these gyros into the real world? And that is really intense and like geni can't solve that problem. Yeah, it's interesting. I think you're you reminded me of the conversations we've had in the past about AI explanability as well. Um, what is that like in the realm of robotics? Like would you say that it sounds to me like it's things are more explainable given everything we've just said about how the software works for robots, but you know, would you say that's something that robotics kind of has an edge over on Gener as well?
I think so. But I think in the sense that yeah like a lot of these systems like they're quite defined in that they have an input they have an output and the model is known to us. So like I think if something goes wrong it's typically quite explainable why it went wrong. Like even you can explain this with Teslas when there's a crash like it typically knows everything that happened up until the crash why it did a certain response and you can kind of work out because again it's quite intuitive and there's only a few options and a few inputs you can work out why it happened but then there's always like an unknown unknown right you always have this parameter that goes into all robotics that's like noise and it's like there will always be some noise and you try to minimize that noise as much as possible by measuring it as much as possible and training as much as possible but there will always be some noise but yeah in most context like
it is explainable Yeah. So, I guess like in the Tesla example, you're basically saying, well, we know what the model looks like and we know I was driving at this speed and the clutch pedal was, you know, at at this depression and the brakes were applied at this time and you can work out, you know, the latencies of everything and then you can work out why something went wrong basically. Whereas, yeah, in the in the kind of Genai case, the chat GPT, how do you work out, you know, how it arrived at this summary of a book?
Well, it yeah, good luck because it's based on, you know, the entire internet and everything they've said. So, that makes sense. Um, exactly. I think that this space is like it's because people are seeing what, you know, LLMs can do now and what GI can do now. They assume that that is intelligence and they assume that like, you know, writing a a 20page report on something very technical is a very hard to do. That's how we define intelligence, right? So like because LLMs can do this very easily now or we think they can do it very well and very easily.
We assume assume that these systems are now extremely intelligent. Most people assume that and we think of walking as very easy and intuitive. We think of opening a door as very easy and intuitive. Recognizing an apple was easy and intuitive. So they assume that these systems now I prompt a question gives me a response can now do those very simple tasks as well if they're plugged into a robot. But they can't because those things that we think of as very easy and intuitive are actually incredibly hard.
And like these systems are just simulating intelligence and just understanding patterns of language and producing output on that. And when you take it to a real defined problem like a robotic system with multiple actuators, multiple control systems, variable dynamic environments, you can't fake it. Like that has to work and it has to work very quickly or you know very quickly and very accurately or things go wrong. And that is still a very hard problem to solve. Yeah, that's really interesting because it's like I think people underestimate just how much complexity there is in the human body for like catching a ball when it's thrown at you, right?
Like how much you how much uh stimulus, your processing, the reaction times, coordination of all the muscles and everything in your body to go and actually catch that ball. Yeah, that's that's something that's not helped by a big large language model. That's something that requires all this other stuff on that's like the bridging of of the AI the classical AI systems with the mechanical right which which seems like where the big challenge is that robotics is exactly and it's why like we've it's taken so long to go to humanoid robots because there's just so many different um parameters involved.
There's so many different joints, so many different like control systems like complex control systems required. And when we're gonna so I think like you know we we're going to talk about the future of robotics and what it will look like and what humanoid robots might be like in the future. You know spoiler alert humanoid. Um but the reality is they're already here and like when you look at a lot of the deployments in sectors today and it is widespread they are far more simplistic than most people think because you know the intelligence is like it's actually really hard to map that intelligence to actuation and controls in dynamic environments.
So you always want to start simple. All right. Well, I've really enjoyed the uh the little schooling lesson I've had here, Alec, from you because it's been it's been really interesting to learn from someone, you know, with with a PhD in this stuff that is uh is that is on the ground. Um my last kind of question on this, I think, before we wrap things up, right now we've covered the five big components is at what point would Azimov's rules come in here? So, where is this going to be in the software part?
So, just as a reminder from the last episode, uh those are the kind of three rules from Azimov. So a robot must not harm a human. A robot must obey commands unless it conflicts with harming a human. And a robot um must not harm itself unless it conflicts with the the former two rules. So at what point do those do these just get programmed in or is this something more complex again? Well, like this is it's a tricky one as well. So in again Isaac Asimov wrote that a hundred years ago basically and it was so ridiculous at the time and now it's becoming more of a reality.
And in that sci-fi novel, I think it's quite interesting to talk about it. the hardware and the software is intrinsically linked. Like they created these systems so that these posatronic ba brains these like the hardware there was also the software as well for these humanoid robots. You couldn't build one of these things without the software deeply linked into it. So it was kind of like hardware and software at the same time. But this is like a really interesting question in robotics where it's like that doesn't have to happen, right?
Like a lot of these systems will be implemented in the software for simplicity. But the question is like okay as these kind of move into the world and like become right now they're still operating in very controlled environments with specialists around them but as they kind of hit that retail adoption where they are operating in in the world there's kind of like I guess one option you could do is to make them so weak that they couldn't possibly hurt someone right that's like what they've done with a lot of these Tesla robots and whatnot but when you have something like a car like you know a great example is all the Teslas when you know they're forced to crash into two two people like either crashing the person in front of them or the person behind them.
Do they go for like an iRoot style? Okay, who's the person who's got the higher likelihood of surviving? Do I hit the old person over here or the two children over here? They're not at that stage, right? They're just at the moment, they're just at the stage where I need to minimize impact to myself and the people in my car, for example. But it is becoming a real question of how are those assessments made? Because as they become more advanced, as they're taking in more sensory data, can handle more compute at the time, they could start to take those factors into account and be like, "Okay, there's a family behind me, there's an individual in front of me.
I'd rather plow into the person in front of me than damage all four people behind me." And it's an ethical concern there. Like, does the person in front take that risk? And or do they sue the, you know, Tesla and the car for the Tesla intentionally accelerating to avoid the hurt to the family behind? Those ethical dilemas are they're for the programmers, right? They're for the companies to set up. And it's really interesting to see how it plays out. Yeah, that that's that's really interesting, right?
It's like you're totally right. It's completely it's down to the the trolley problems, right? You're going to have all these ethical trolley problems where your your computer is deciding how to minimize uh suffering essentially, human suffering in these cases in accordance with the first law. Um, and something I hadn't even thought about until you started talking about Teslas then in this context. But maybe in the future there'd also be like a corporate conflict there because maybe, you know, if you're if you're in the case of a electronic car manufacturer, then it's not very appealing if you know your car that I've the car that I've bought is is likely to minimize suffering and it doesn't prioritize me as the person in the car, right?
So maybe maybe the car manufacturers will build that in as a rule and that will conflict with all sorts of things and regulations, but maybe they're happy to be sued and and you know that maybe that's that's more costefficient for them than the lack of sales they get if people knew that the cars would might actually be you know not prioritizing their own safety. And this is like the beautiful part about all this stuff related to AI and robotics and humanoid robotics is it really brings in as me and Jack love philosophy into this.
It's like how do you make these decisions? And eventually someone is going to have to start thinking about it and programming these in. Right now they're very simplistic machines where it's just like you know basically shut down if anything goes wrong or that style but more and more we're going to have to have these ethical dilemma questions as well about what should be done in a certain environment. And you can apply this to you know militarygrade robotics as well like how are they making the decision to kill people for example.
And for those that have read ice if you haven't you definitely should. They go into this into a a huge amount of detail. It's really interesting, insightful to be like, what is pain for a human? Like you can have emotional pain like how do you measure pain? There's no real objective way to measure it. So it's like really difficult to say, you know, pain versus this person versus that person. If I have two options, which one do I go for? And yeah, I think it's maybe a bit beyond us right now, but maybe in 20 years, like these will be the major questions that we're asking.
But I think for now, maybe we should call the second part of this three-part series to an end. Again, I love this. I'm so glad we're finally doing robotics. Um, so in this episode, we broke down, you know, what makes robots tick. The controllers that think, the sensors that perceive, the motors that move, the power systems that keep it running, and probably most interestingly, the software that ties it all together. Um, but yeah, thanks for listening to this episode. Join us next time for part three where we'll talk about how robots are being used right now and we'll hint to the future of robotics, what trends are emerging, and what our world could look like in 50 years time when we have humanoid robots absolutely everywhere.
Thank you for listening wherever you may be and join us next time to untangle a little more of web
Topics: AI & Robotics · Blockchain & Crypto