Humanoids, How Far To Go_

Okay, so here's a number to start with. Every single week, Waymo says, it gives more than half a million rides with nobody in the driver's seat in more than 10 American cities. And it's already lining up London and Tokyo next. Now here's another number. The best-known humanoid robot factory trial, Figure Robot, at a BMW plant in the U.S., ran for about 1,250 hours, Figure says, in total, over 11 months. So one kind of robot is already a business, and the other is still mostly a pilot. And yet the money going into humanoids is huge, with Figure alone valued at $39 billion. Self-driving cars took about 20 years to get from a desert race to a paid ride. So what will it take for the robot that's meant to work right next to you, in a factory or a warehouse or your kitchen? I think the best way to answer that is to look at what cars had to get through first. Because there were three gates. The problem had to be narrow enough. The machine had to be provably safe. And it had to get cheap. Let's take them one at a time. 2004, the Mojave Desert. DARPA, the Pentagon's research agency, offers a million dollars to any robot vehicle that can drive 142 miles on its own, from Barstow in California to Prim in Nevada. Nobody finishes. The best car manages about seven miles, and then it gets stuck on a rock. That was it, the state of the art, about 20 years ago. A field full of serious robotics teams, a million-dollar prize on the table, and not one car could get more than a few miles down a desert road. The next year, the race runs again, and this time, five vehicles make it all the way to the end. Stanford's car, Stanley, wins in just under seven hours and takes home $2 million. Then, in 2007, the robots move into traffic with other cars on the road, both robot and human-driven. And Carnegie Mellon's car, Boss, wins that one. And here's the part that matters for our story. In 2009, Google quietly starts its own self-driving project. And it hires, in its own words, the very best engineers from the DARPA challenges. Its first goal is wonderfully specific. Ten routes around California, each about 100 miles long, driven with no human help. And it ticks off the last one in 2010. By 2015, one of its cars drives a man called Steve Mahan around Austin, on ordinary streets, with no test driver and no police escort. Steve is legally blind, and the Waymo engineer who organized the ride wrote that it was his first time alone in a car in 12 years. A year later, in 2016, the project becomes Waymo. So, Desert Race to Public Road took about 11 years, and Public Road to Half a Million Rides a Week took about another 10. That's the 20 years. And it isn't just Waymo now, either. Baidu's Apollo Go runs in 28 cities, and Amazon's Zoox started charging for rides in Las Vegas this August. Now, why did cars get there at all? Because driving, for all its danger, is a narrow problem. Think about what a car actually has to do. It stays in a lane, it follows rules that everybody already agreed on, and in the end, it's choosing between steering, speeding up, and slowing down, which is a pretty small set of outputs. Back in 2016, NVIDIA showed that a car could learn to steer straight from the pixels of a single camera with very little human training data. A humanoid is a different animal. Boston Dynamics' production Atlas has 56 degrees of freedom, which means 56 separate ways to move, all controlled at once while the robot keeps its balance. And then there are hands. One human hand has about 17,000 touch receptors, around 1,000 in each fingertip, while most robots are still working mainly from cameras. There's an old idea in robotics called Moravec's paradox. Back in 1988, Hans Moravec pointed out that it's comparatively easy to make a computer do adult things, like pass a test, and really hard to give it the perception and movement of a toddler. So, cars got the easy version of the hard problem, and humanoids got the hard version. The second thing cars had was data, and lots of it, because every Waymo is also a data collector. Waymo says its cars have driven more than 270 million miles with nobody behind the wheel, and that they add about 4 million more every week. Tesla says drivers using its supervised driving software are approaching 12 billion miles. Robots have nothing like that. The biggest shared open robot dataset, a project called OpenX Embodiment, pooled about a million robot demonstrations. Which sounds like a lot, until you notice it's spread across 22 different robot bodies and more than 500 skills. A startup called Generalist says it now has 270,000 hours of real-world handling data, growing by 10,000 hours a week. That's the company's own claim, and it's the biggest anyone has made in public. Miles, hours, and demonstrations aren't the same unit, so don't line them up directly. But the gap is pretty obvious. So how are humanoid makers getting through gate one? They're cheating a little — in a smart way — by shrinking the problem. Instead of a robot that can do anything, you get a robot that does one job in one place. Agility's Digit moves plastic totes onto conveyors in a warehouse. Figure's robot at BMW loaded sheet metal parts, and the company says it helped build more than 30,000 cars. And Hyundai plans to put Atlas into its car plants from 2028, sorting parts first and adding assembly work by 2030. The brains are starting to look alike, too. Figure, Google DeepMind, and NVIDIA all describe the same basic design. A slower model that looks at the scene and plans, sitting on top of a faster one that moves the joints. In Figure's case, the planner runs about seven to nine times a second, and the mover runs 200 times a second. So gate one is being passed, just slowly and one task at a time. Now, gate two, safety. And this one has a much darker history. On the 25th of January, 1979, A 25-year-old worker named Robert Williams was killed at Ford's casting plant in Flat Rock, Michigan when he was hit by the arm of a robotic parts retrieval system. He's believed to be the first person ever killed by a robot. Two years later in Japan, a maintenance worker at Kawasaki named Kenji Urada climbed over a fence to get to a robot, and it crushed him. The fence was built to cut the robot's power when it was opened, but he went over it instead. Industrial robots weren't new, even then. The first Unimate went to work back in 1961 at a General Motors plant in New Jersey, unloading parts from a die-casting press. And America's Workplace Safety Regulator, OSHA, still notes that a lot of robot accidents don't happen during normal work at all, but during setup, programming, and maintenance, when a person is inside the robot's space. Factory robots have long followed one simple rule. Keep people out. Fences, locked gates, light curtains, and pressure mats. And if a person gets inside the robot's space, the power goes off. That's still how most factory robots work. And the main international standard for them, ISO 10218, got its first big rewrite since 2011, only last year. Then came collaborative robots, slower arms with limits on how hard they're allowed to push so that people could work beside them. Here's the problem, a humanoid breaks both halves of that rule. It's meant to work right next to you, so you can't fence it off. And you can't just cut the power either, because the moment you do, it falls over. ISO's own draft standard describes these machines, quite literally, as robots that could fall over without power. That standard, by the way, isn't finished. Work started in May 2025. The comment period on the draft only closed this July. And it's still a draft. So right now, humanoids are being built faster than the rules for them. Companies are working around it. Agility's answer is something called a Category 1 stop, where the motors stay powered while the robot slows down. And only then is the power cut. Its newest Digit, launched in September, goes a step further. And Agility says that when it senses a person coming close, it can change direction, stop, or simply sit down and switch off. And Agility's co-founder has a refreshingly honest take on falling. Just fall. You'll get back up. Now compare that with cars. Robo-taxis didn't get permission because people trusted them. They got it by piling up evidence. Waymo says it has driven those 270 million miles with nobody at the wheel. And that, compared with human drivers on the same roads, it's had about 95% fewer crashes causing serious injury. That's Waymo's own number. But Waymo's researchers have also put their data through peer-reviewed journals. And in one study with the insurer Swiss Re, They compared 25 million Waymo miles with insurance claims and found two injury claims, where they'd expect about 26 for human drivers. And regulators watched every crash. Since 2021, America's road safety regulator has required companies to report crashes that involve automated driving. So the evidence was out in the open. In August 2023, California's utilities regulator voted 3-1 to let Waymo and Cruise charge fares for driverless rides in San Francisco at any hour of the day. And that's the moment robo-taxis became a real business there. The one time a company broke that trust, it ended badly. In October 2023, in San Francisco, a hit-and-run driver knocked a pedestrian into the path of a Cruise robo-taxi. The car stopped and then, not sensing she was underneath, it dragged her about six meters. Within weeks, California suspended Cruise's permits, saying the company had misrepresented what happened. So it wasn't only the crash, it was the misleading account of it that cost Cruise its permits, and then fines, including a criminal one for a false report to the federal regulator. And in December 2024, GM shut its robo-taxi business down. Which brings us to the really hard bit. How do you certify a robot that's run by a learned AI model? Normal safety engineering starts by writing down exactly what a system must do, and then testing the code against that list. With machine learning, as the safety researcher Philip Koopman puts it, you don't know how to write the requirements. The car world's answer has been a stack of standards and a trick. You don't try to certify the AI itself. You wrap it in a simpler, old-fashioned safety layer that you can certify, and that's allowed to overrule it. And an international technical report from 2024 describes exactly that pattern. So, in a funny way, the cage is back. It's just made of code now. Then there's the home, with no fences, no trained staff, and kids and dogs everywhere. The ISO draft is written for industry, not for your living room. So companies are designing for softness instead. One X says its Neo weighs about 30 kilos and has a soft polymer body with no pinch points. And Figure says its newest robot is wrapped in soft textile covers, with foam where fingers could get caught. And safety is already a legal fight. A former head of product safety at Figure is suing the company. He alleges that he warned executives its robots were powerful enough to fracture a human skull, and that he was fired days after his strongest complaints. Figure says he was let go for poor performance, calls his claims falsehoods, and has filed counterclaims. And as of July, the case was still before the court. His lawyer says it may be among the first whistleblower cases about humanoid safety. So, gate two. For cars, it took hundreds of millions of public miles, and regulators reading every crash report. For humanoids, we couldn't find a single maker that publishes how often its robots need a human to step in. So that count hasn't really started. OK, gate three, price. And this is where the car story gets really useful. Back in the DARPA days, the spinning laser scanner on top of many of the robot cars, a Velodyne LiDAR, cost about $75,000. By 2014, Velodyne was selling a smaller one for about $8,000. And Waymo built its own and said it cut the cost by more than 90%. And last year, a Chinese maker, Hesai, put a LiDAR into mass production at about $200. So that's $75,000 down to $200 in 18 years. They're not exactly the same kind of sensor, but you get the idea. And here's the twist. Cheap LiDAR didn't save every LiDAR maker. Luminar, once one of the stars, filed for bankruptcy after losing its contract with Volvo. Cheap is great for buyers. It can be brutal for sellers. So what's a humanoid made of? And why does it cost so much? Mostly, joints. Bank of America's reference design has 16 rotating joints and 14 that push in a straight line, plus a dexterous hand. And the joints alone can be more than 30% of the parts bill. The computer is one line among many, and NVIDIA's Robot Brain Developer Kit launched at about $3,500, though it's now closer to $5,500. Those costs are falling fast. Goldman Sachs found that the cost to build a humanoid dropped by about 40% in a single year, from somewhere between $50,000 and $250,000 to between $30,000 and $150,000. And it's cheapest in China. Morgan Stanley estimated that the same Tesla Optimus would cost about $46,000 to build with Chinese parts, and about $131,000 without them. By Morgan Stanley's count, roughly nine in ten humanoids shipped last year came from Chinese companies. And UBTECH alone sold 921 full-size humanoids in the first half of this year. Look at Unitree. Its own listing documents show the average price of its humanoids fell from about 590,000 yuan in 2023 to about 168,000 in the first nine months of 2025. That's a drop of more than 70 percent, and yet its gross margin stayed around 60 percent. Unitree says it shipped more than five and a half thousand humanoids last year, which would make it the biggest maker, though one research firm puts its rival Agibot ahead. You can buy its small G1 for about $13,500 or its full-size H2 for about $40,000. But who's actually buying them? Unitree's filing says about three-quarters of its humanoid revenue came from research and education buyers. Industrial customers were about 9%. And a big chunk of that slice was robots giving guided tours. So, cheap? Yes. Doing factory work? Not really. And not yet. The American makers are scaling up, too, just from a smaller base. Figure says its factory went from one robot a day to one an hour, and that it has delivered more than 350 of its newest robot, though it doesn't say how many went to paying customers. Then there's the home. 1X will sell you its Neo for $20,000 or rent it to you for $499 a month. But a Wall Street Journal reporter who tested it found that a person in a VR headset was running every task. 1X is open about this. It calls it expert mode, and owners have to agree to it. 1X says it has been shipping to early buyers since the spring. But as of mid-October, we couldn't find an independent report of one actually arriving in a customer's home. For businesses, the model is renting. Agility's filing shows it rents Digit for about $8,500 a month, plus a $25,000 setup fee. So let's do the maths. On one 8-hour shift a day, that comes to roughly $50 an hour. And the median pay for a U.S. warehouse worker who moves stock around is about $18 an hour. So Digit really only competes if it works two or three shifts a day. — And Agility's revenue last year was $1.8 million, against an operating loss of $140 million. To be fair, robo-taxis aren't making easy money either. Alphabet's Other Bets, the part of the company that's mostly Waymo, lost $1.8 billion in a single quarter this year. So cheap enough to sell is not the same as profitable. And the banks aren't short of optimism. In 2024, Goldman Sachs forecast a $38 billion humanoid market by 2035. And this year, it reportedly more than tripled that. Morgan Stanley talks about $5 trillion by 2050. Those are models, not facts. And like most forecasts, they keep moving. So let's go back to the three gates. Gate one, a narrow enough problem. Humanoids are getting through, one job at a time. Moving totes, loading sheet metal, sorting parts. That's real progress. And it's how cars did it too, one city at a time. Gate two, proof of safety. This is the one I'd watch most closely, because the standard isn't finished. Nobody publishes how often their robots need help, and a safety whistleblower case against Figure, which Figure denies, was still before the court as of July. Cars got through this gate with hundreds of millions of public miles. And humanoids haven't started that count. Gate 3, cost. It's moving fastest, mostly because of China. And prices are falling the way LIDARs did. But the cheap robots are mostly going to labs and universities. The useful ones are still expensive. And big names on both sides of the Pacific, like Agility and Ubitech, are still losing money. So how long? Tesla talked about building 5,000 to 10,000 Optimus robots in 2025, and the tech site Electrek found no evidence it built even a few hundred. Hyundai's own date for Atlas on its lines is 2028. And the roboticist Rodney Brooks says we're more than 10 years away from the first profitable deployment of humanoids. If I had to bet, I'd say humanoids are roughly where self-driving cars were about 15 years ago. Past the desert race, but not yet at the paid ride. The money says it's close. The evidence says it's early. And the gap between those two is the thing to watch. That's it for this one. Thanks for watching, and thank you for listening.