The Great Humanoid Robot Hoax

Why are we obsessed with building robots that look like us? Boston Dynamics revealed their first humanoid robot 12 years ago, and it looked like it was just a matter of time until you had one of these pouring your coffee, right? But what has been poured is money. Billions of dollars, not just into Boston Dynamics, but into dozens of companies that have yet to deliver on the promise. It almost leaves you wondering if humanoid robots are just viral video machines rather than productivity machines. And when you look at this chart, the answer becomes almost too obvious, but we'll get back to this in a moment. Let's get something out of the way first. We humans, we're not the most efficient machines. Starting with the basics, how much energy do we need to move? Walking 100 meters or 300 feet in a straight line would consume about 30 kilojoules or about seven calories. That sounds like nothing, and it mostly is nothing. It'd be about 14 grams of an apple. And the fact that our bodies can convert apples to steps and then into YouTube videos is an evolution marvel, but that's not the point of this video. Walking that distance requires some pretty inefficient locomotion. I have to lift each leg from the ground, accelerate it from a still position into movement, and then stop probably around 200 times. On the contrary, if I could just roll that distance, I'd only really need about 3.4 kilojoules of energy or one calorie, assuming a flat surface. The real complicated part about walking is that my muscles need to squeeze and stretch, my ears need to keep my balance, and my entire nervous system has to coordinate these micro adjustments in real time. Every step requires signals between my brain, spinal cord, muscles, tendons, and joints. My eyes need to feed me visual cues about where to place my feet, while my vestibular system in my inner ear monitors my head position and motion to keep me upright. Muscle spindles detect changes in muscle length, helping fine-tune my posture so I don't trip or wobble. Why did we feel the need to reverse-engineer this whole mess when we could just use this instead? To do all this, a robot needs cameras, lighter force sensors, machine learning for motion planning. Even routine tasks like lifting an arm or bending a knee can draw hundreds of watts, both for the movement and for the processing power, which means that most of modern robots, if operating on battery, only have a few hours of autonomy. Moreover, because of all the work needed to maintain this balance, most robots can only lift about 30kg or 65lbs, which is not much more than a skinny human. So why take that anthropomorphism approach? Why not just build an efficient machine? Is it a lack of imagination? Maybe. And part of it is definitely that we just like things that look like us. We humans are just stubborn about giving human shapes to things. The Greeks imagined Hephaestus, this god of blacksmiths, that created a giant bronze automaton, Talos, to guard the island of Crete. A Chinese text from the 3rd century BCE tells this tale of Yang Shi, an engineer that built a life-size, human-like robot for King Mu, made of leather and wood that could supposedly sing. Though technically they weren't really called robots until the 1900s when this Czech playwright coined the name in his play RUR, Rossum's Universal Robots. I think Hollywood production designers have shaped more of our lives than we'd like to let on. And while having robots shaped like humans was an easy way to cut some production costs for a movie, in the real world it's been nothing but a headache. One of the arguments for humanoid shapes is connection. A humanoid robot is supposed to be more relatable, to create a connection, to make them understandable, instead of alien, in theory. But we discovered something along the way. Japanese roboticist Masahino Mori coined the term in 1970 using this two-axis chart. On the vertical axis, he used the Japanese word which doesn't really translate to English very well. Might mean something like affinity, familiarity, report, likeability, if you will. And then he observed that a toy robot, for example, would create more shinwakan than an industrial robot because it was further down the right in this horizontal axis, which is human-likeness. However, as you get closer to 100% human-likeness, the inverse happens. Robots start creating these negative reactions, eagerness, discomfort, and even repulsion. Now, there's an evolutionary trait that gets triggered as we reach this kind of area of the chart, which is our natural, instinctive ability to detect unhealthy members of our species. It's a really fucked up feature that makes us naturally discriminate against people, for example, who have disabilities. But all research points to this being this evolution trait that helps us find suitable mates. Mori called this Bukimi no Tani, or the uncanny valley, which has been the subject of hundreds of studies since and confirmed as true. Staying on the left side of that valley, as a safe robot-looking robot, or venturing to try and cross this valley is no small dilemma for these companies, but that's not the biggest one by a mile. Despite all of those hurdles and inefficiencies in our anatomy, scientists have been able to When there's a will and millions of dollars in R&D, there is a way. And the latest humanoid prototypes have overcome our bipedal limitations and, dare I say, even improved on them. But we've wasted all of this time teaching them to walk. We haven't even gotten to teaching them how to, well, how to do anything else. What kind of shoes are you wearing? Are those Skechers? Now, it's easy to be fooled by the viral videos, you know, because when you see a robot vacuuming a carpet, or another one kung fu fighting, or another one pouring you coffee on top of this humanoid machine, like, it looks like we're there. Like, it looks like we got there. Like, you could buy one tomorrow. It's funny, because when you Google most of these companies, their descriptions are always the same. Oh, we're the first humanoid robot to do, or we're the first general purpose robot. But when you search for robots, none of these companies pop up in the front page. Now, the first one is Boston Dynamics, which was down on the second page for me. And yet, Boston Dynamics gets about 100,000 visits per month on its website, mostly from people searching for the name of the brand. Now, these companies' strategy for fundraising, recruiting, selling, is purely PR from their breakthroughs. But if you don't have a robot that makes viral videos, you should probably pay more attention to those Google searches, and our sponsor for today can help you with that. Now, no kidding, here's my actual subscription history for Ahrefs. We've been their customers since 2015, and they've been a key platform for improving our SEO, monitoring competition, identifying which keywords they're trying to snatch us, and finding keyword ideas for new content. Now, they also recently launched Ahrefs Analytics. If you've been to Google Analytics lately, it's really an utter mess to use. And Ahrefs lets you get traffic reports across countries and devices without any of the hassle. It's privacy-friendly and requires no cookies and no personal data. And it's also real-time, no more of that waiting till the next day for your reports. You can set it up in seconds, and it's completely free. You can just use the link in the description or snap this QR code here. Really, Ahrefs is my golden standard for what an SEO tool should do. Anyway, despite the hype, none of these companies has actually offered a release date, and most experts just question their ability to ever ship a robot for under $20,000. One of the biggest problems is that none of them can really do anything. Most of the videos you've seen were shot in highly controlled environments, with some editing magic to skip the bloopers, and with robots performing simple, pre-programmed tasks. Elon knows how to play stunts to bump his stock numbers, so he's showing his Tesla bot, Optimum, doing all sorts of things. But multiple reports have exposed that employees were partly controlling them, so it's really hard to say what they were actually doing by themselves. Aptronix Apollo is supposed to be a general-purpose robot, but has not been used outside of a couple of limited pilots. Xiaomi has CyberOne, but there are also no live tests. I had my hopes up for Uptek Robotics. They went public, they have revenue, they did this control demo six years ago with their Walker robot, so they gotta have something, right? Well, yeah, they make revenue, but none of their commercial robots are humanoid. Boston Dynamics' Atlas can even breakdance now. They are THE robotics company. They gotta have something. Well, yeah, they do make revenue from their Stretch and their Spot robots, but Atlas has never really left the lab. The only ones that are slightly, slightly ahead are Agility Robotics' Digit, which moved some boxes in a warehouse in a very limited test, and Figure 2, from a company called Figure AI, which kinda worked in a BMW assembly line for a week. And those are the only two viable uses that these companies have been pushing, and the only ones that we have some visibility that they could potentially accomplish soon. Boxes in a conveyor belt and part of an assembly line in a car factory. Uses that already have robots purposely built for those tasks. Do we need billions of dollars in R&D for a human-shaped version of either of these? The most crucial challenge that we need to overcome in robots is a problem called Moravec's Paradox, which basically says that it's comparatively easy to make computers exhibit adult-level performance on intelligence tests and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility. Now, let me break that into simple words. Take this simple math operation. It should have taken you, like, a second to solve, and it takes a computer a fraction of a processing cycle to solve. It's numbers, adding symbols, it's the language that computers speak. If you give this same operation to a toddler, they probably wouldn't be able to figure it out until they understand counting, until they understand the concept of addition, until they can read numbers and math. I know so much more than babies do, and it can be hard to forget all that stuff. Now, with the current state of generative AI, it's even easy for a computer to find a fun joke reference for this, even if it can't really laugh at it. One thing a toddler can do, with little difficulty, is to solve this thing. And solving this thing for a robot, it's extremely, extremely hard. The colors and the shapes on this thing are almost inviting for a human to touch and to explore. It's curiosity, right? And reaching out and grabbing it doesn't require us to think much. We naturally infer the weight and the force that we need to apply to lift each piece, from years of experiencing lifting similar objects. After we're about one year old, we have enough information to know which piece fits where. But to a robot, this is completely different. For starters, a robot doesn't even have curiosity. Some instructions need to be written for it so that it tells what to do. Maybe an instruction like play. Well, play is still too abstract of a word. We need to break it down, okay? So, understand game still requires a lot of work. It's seeing, it's grabbing, it's testing. And each one of them is massively complex. Seeing, it's not just booting a camera. It's understanding what's out there. It's understanding 3D space, distance, size. And the processor inside this robot doesn't see an image, really. It's just seeing bits. To a robot, each pixel is just a number. It must first group these pixels together, understand color, what represents an edge, what represents depth. But even identifying color is tricky. Colors can change based on lighting, shadows, reflections. Humans just adjust very quickly to this, and robots don't. We've made strides in image recognition these past couple years. Current AI models can help solve that, but grabbing is harder. These thousands of tiny details that are so instinctive for us require explicit instructions, code, complex calculations, careful programming. Actions that we do without thinking are incredibly hard for robots, no matter how smart they seem. That's more of xParadox. Doing all that stuff is the result of millions of years of evolution that are embedded in our code. Stray dogs are pretty good at crossing roads, and they didn't need a product roadmap or a Series B funding to do it. Because there's an evolutionary trait that has been passed down through generations. And as we start building these machines, we haven't even begun to codify all of this knowledge. Which means that, even today, we're many years away from having a robot solve existing in the world. But robots can still be so useful. This is the Unimate, which is the first ever proper robot installed in a car assembly line. 60 years ago. And the whole point of it was that it could do things that humans couldn't. People didn't love it at the time. It led to cheaper cars. But everyone was worried that it could replace humans at the assembly lines. And they were right. But imagine the PR nightmare if these car companies start firing employees this year to replace them with these human-shaped robots. Something about the human shape makes it sting more, doesn't it? But back to the companies making them. What is the endgame? Do these companies and their investors realistically see a path to ever turning a profit with these things? Again, the answer might be in this chart. So, let me run some numbers for you. That massive spike there is 2025. And that's just the first quarter of 2025, led by Figure.ai's $1.5 billion round. So, let's dig on this example. According to a Reuters report, as we're shooting this today, end of March, the company is on track to raise a Series C round of $1.5 billion on a $39.5 billion valuation. That would bring the total amount of funding for this company to $2 billion, all of which, we can assume, is going to be spent on R&D. Now, considering the funding amount and the valuation, we can conclude that these investors are going to end up with about 3.6% of the company. Now, how do you get that money back? Will you ever get it back? Let's assume that this company will pay about 50% of its profits as dividends, which is a reasonably usual standard. That means that the company needs $3 billion in net income, after taxes, to get investors to break even. Companies in this space tend to operate on net income margins of about 10%. Let's be generous. Let's assume that this company will hit 15% margins, because it's the first one to market, the first one with a functional robot. That would mean that this company needs to make about $83 billion in net income to extract the $1.5 billion that these investors would get, based on their 3% ownership. Now, we're looking at a company that maybe needs to make about $550 billion in revenue to extract enough money to pay investors back. That's more than Apple's revenue and about nine times NVIDIA's total revenue for 2024. Now, let's say that they manage to sell these things for $50,000 per robot. So, we're looking at 11 million robots sold. For example, today, there are 7 million warehouse workers worldwide. They need to replace them all. So, how many years will it take for this company to make this revenue back? Will it ever make it back? How many jobs would these robots have to replace to get this company to positive ROI? Of course, most investors aren't just going to sit around and wait for profits. There's a much faster way to realize these gains. What if, instead of waiting on years of profit margins, the company just hits the stock market? It doesn't matter if it's profitable or not. We know that now. So, anybody can invest in the future of robotics. Now, retail investors, speculators, could they pay enough money per share to bring this company to a market cap of over $40 billion, get these investors their money back? Look at Rivian, for example. Rivian was once worth $150 billion market cap. Back then, they had just shipped about 200 cars. Hold that thought for a sec. Now, let's look at a real example. What happened with Tesla, which was already public, when Optimum was announced? Tesla's stock price shot up 55% after Optimum was announced, creating about $450 billion in value practically overnight, just because Tesla announced that they were pursuing robots. On the other hand, Boston Dynamics has been passed as a hot potato between Google SoftBank, Hyundai, with valuations around a billion dollars, justified not by its profits, hardly by its sales, but mostly because of its IP. At least right now, robotics companies aren't selling robots. They're selling hype. And hype scales faster than reality. Can you do this? I love you. You know how to do that? Now, don't get me wrong. I'm not trying to say that this is a hoax. Like, using the Boston Dynamics example, while Atlas isn't hitting store shelves anytime soon, the innovations, tech, patents, IP, developed to train, have been incredibly useful to develop other technologies for other robots that they manufacture and make money from. Probably the most important breakthrough we've made to solve this eventually is through virtual environments. One of the reasons why our current generative AI models have gotten so good at replicating how we write, at generating images that look real, is because it's had billions of texts and millions of images to train from. But there's very little training data for robots. There's no camera inside of a baby's eye recording how he fumbles with his toys and learns to solve a puzzle. There's no data on how to navigate a kitchen or to figure out the most likely drawer where the owner keeps the spoons. And creating training data for this stuff is very hard and very expensive. But there's hope. And it comes in the form of virtual worlds. It's a very hot topic these days. NVIDIA has developed three technologies to create a framework for this. IsaacSim, which provides the simulation environment. Cosmos, which supplies the synthetic training data. And Newton, that ensures accurate physics modeling. My man Devrim started a company called Lucky Robots precisely to solve this thing. To create these environments where robots can train on. And this wasn't even possible a few years ago because the resolution at which these environments rendered wasn't enough to resemble the real world. So it was useless for robots cameras to train. There are very few technologies where we've been able to pour so many billions of dollars and still yielded profitable results. Nuclear fusion is still a question mark. We don't know if we're going to solve that. So is quantum computing. Space exploration is probably one where there is at least some ROI, or at least for Elon. And AI or generative AI is a bigger question, especially if we are not able to get around this equation, which is the subject of our video from a couple of weeks back. Check it out. See you in the next one.