The Future of Humanoid Robotics | Jonathan Hurst | TEDxPortland

Ajouté : 2026-06-04

[00:00:26]All right.

[00:00:31]Robots are arriving in a really big way.

[00:00:35]And robots aren't entirely new.

[00:00:38]Certainly, the first era of robotics, I would call it, started in the 60s with industrial robot arms and manufacturing and the automotive industry. Maybe a second era of robotics started in 2000s uh with mobile robots, the kinds that you see in an Amazon warehouse rolling around carrying things throughout throughout the entire facility. But what's really new right now and what's different is AI enabling these machines to be more general and be more broad and be able to perceive the their things in their world and decide what to do with them. and um applying them to robots is really making them into an embodied AI.

[00:01:17]And this is new. This hasn't been done before.

[00:01:21]Uh analysts are starting to look at this and predict how are these going to impact our lives? How is it going to impact a market and how big are these markets going to be just for the robots?

[00:01:33]Uh which of course are going to be in all kinds of other um influencing all kinds of other markets.

[00:01:39]And they're predicting that by 2050 the market for humanoids, which is not all of um embodied AI, but just the humanoids, is going to be twice the size of the automotive market today.

[00:01:51]So when you think about how often we use our smartphones and how much cars are part of our lives and then imagine a market that's twice that size for humanoid robots and you can start to get an idea of just how much people think these are going to be part of everyday human life and really influencing us.

[00:02:12]Now, there's a couple of big ways that I believe AI, embodied AI, humanoid robots and others are going to influence our economy and influence jobs and ultimately then our quality of life. One is capacity. These robots are going to be able to fill these massive labor gaps and do things for which we just don't have enough people to do today. And this is a lot like think about American history. At one point, 90% of Americans worked in agriculture.

[00:02:41]And today only about 2% of Americans work in agriculture.

[00:02:46]But of course 88% of Americans are not unemployed. We just have most people in the United States doing all kinds of tasks and roles and producing things that we never could have imagined when 90% of Americans worked in agriculture.

[00:03:00]We build so much more than we did at that point in time. And I expect that trend to continue as we move forward and have new forms of automation, which by the way, the reason we can do this is because of machinery, because of laborsaving devices, humanoid robots are going to be like that. We'll be able to produce a lot more than we can today.

[00:03:20]The other piece though is capability.

[00:03:22]And for a good analogy, a rhetorical question is how many blacksmiths does it take to build a laptop?

[00:03:31]Of course, it doesn't matter how many blacksmiths you have. they can't build a laptop, right? You need this ladder of technologies that enable you to build the tools to build the next thing in order to get to something like a laptop.

[00:03:43]And that's what embodied AI is also going to enable for us. We're going to be able to build things that we cannot create today with human hands. Hard to imagine what that will be. We can all dream and speculate, but that's actually the point is to allow us to dream and speculate big.

[00:03:58]So, I've talked about uh embodied AI a couple times and I thought I would explain that a little bit. Uh most of us have encountered AI in the short in the form of my Google Gemini or or Chat GPT or um Claude's uh Anthropics Claude. But those are really a brain in a box. Those are things that can talk with you with natural language, but they cannot control a physical body of a robot. It doesn't know how to coordinate all the individual joints or learn an individual skill. And in fact, the call it the cognition, call it this uh these large language models. That's kind of the easy AI. Now, I know it's new and I know it's been an incredible effort to create it.

[00:04:39]But it all of this data just exists on the internet. All of language and all of the um things that people have written and all of the images to train these AIs to then interact in that medium in language and in images. But on a robot there is no source of data for how you should coordinate all of the joints to work together. There's no source of that data. So we have to sort of generate that. We have to find a way to create that. I guess I also want to say that AI is a big tent of computational tools.

[00:05:07]It's not just one thing. There's many different tools and different tools are appropriate at different places on the machine like this. So we talked about the cognition side that trained off of all this data. But then skills are something where maybe you would train it using motion capture data or animation.

[00:05:23]uh or teley operation, a person remote controlling a robot or something, some way of showing the robot how to do a thing like grasp a door handle or pick up a mug or some skill that you want to show it how to do. And then the coordination is something that the robot just has to explore itself. That can be done in simulation. And you've heard of reinforcement learning where these machines explore how do you coordinate all of your joints in order to create a certain goal. It needs to explore and figure that out on its own. All of this is predicated on having a machine that's really capable and getting the physics right for the hardware, the physical hardware. And that's good oldfashioned engineering. AI is uh helping engineers with that, but it isn't creating the robots from scratch. Now, there's a really close analogy to how people learn. Think about how you might learn to play a musical instrument, play the violin. The cognition piece means you can talk with people, you can see the violin, you can recognize it, you can get around in the world and recognize the world. Um, the skill is your teacher showing you here's how you hold the instrument. Here's how you hold the bow.

[00:06:20]Here's where you put your fingers. Of course, if that's all you have, you're not going to be a very good violinist yet. You have to go practice for a thousand hours. You have to figure out how your body coordinates well in order to get that sound and to create that behavior that you learned as the skill.

[00:06:34]And all of that is built on being able-bodied, having the right number of fingers, having a good ear, and so on.

[00:06:40]So there are close analogies here um that are emerging as engineers figure out the best ways to apply embodied AI and build machines that can operate in human spaces. So here's a picture of Portland. Um it's what it looks like today. But what if we had this tireless partner to amplify human ambitions to help us really think big? What does the future look like? I don't know. But uh I hope that we will be able to solve some of the big problems. I hope that we will be able to meet all of the basic human needs for all of humanity. I hope we will be able to colonize the solar system. Uh I don't know if we're going to live in glass bubbles, but I know that this is an AI generated image. So if AI is our partner, maybe we will.

[00:07:28]We'll see. Now, in order to dream big or or part of this big dream rather is having robots that can help us in our homes and robots that can do our chores and robots that you can log into remotely and have a teleresence uh you know session with your family members and things like that. Uh but it's going to be a little while. Be patient about this. There's three big reasons. Number one is chaos. Everybody's homes are different. uh an individual's home is different on any given day and the capabilities required of the robot to really navigate well and handle things and do all these chores is is really beyond where robotics is right now. It's going to be a while before robots are going to be able to do useful work in our in our homes. The second one is cost. Families are going to be very costsensitive and these robots are going to have to be very affordable while being exceptionally capable in order to do this. And there's a lot of work that needs to be done, a big flywheel of manufacturing and engineering to get to that point of bringing the cost down to this point. And then finally, the big one, the major blocker is safety. It is completely unacceptable for a robot to fall on your child. And these robots are going to be strong enough to do work, which means they're also strong enough to cause injury. And so you have to figure out how to make this safe in someone's home. Imagine feeling okay with handing your baby to your humanoid robot. I certainly wouldn't. Um, and it'll be a little while before we get to that kind of level of trust of having a robot in our home. Safety is really tough, but we can do this. We can have robots do something relatively straightforward. Pick up bins and move them around. Robots are very capable of doing that kind of thing. Uh, humanoid robots in human spaces. uh things that people were doing, but sort of the classic 3Ds of robotics, the dull, dirty, dangerous things, something repetitive. We can do that. And the value that's generated from these machines doing that covers their cost quite well today.

[00:09:30]And we can make these safe because we have a more structured environment with which to sort of plan out exactly how we're going to make sure that it's safe around people.

[00:09:42]But you know the big question is why don't we have thousands of these deployed right now it really is about the safety. We've just started and Digit is the first commercially deployed humanoid robot.

[00:09:56]This is in a Toyota warehouse. We have it in Amazon. We have it at GXO. We have it at other places. So you can see that Digit can do the work.

[00:10:06]You can also see that there's a physical barrier around these robots.

[00:10:11]that actually separates the robots out from people. So, how do we make these things safe? Right? It isn't about the robot's intent. It isn't about what the robot wants to do or doesn't want to do.

[00:10:21]And that's what a lot of the science fiction is focused on. If you're familiar with Isaac Azimov's three laws of robotics, saying robots may not harm a person. Um, they may not harm themselves and they must obey a person.

[00:10:33]Well, really, you're relying on a robot then to have judgment. And robots don't have judgment. So this is not about a robot's intent. This is about the designer's intent. This is about when you're designing the robot in the environment the robot is going to operate in and understanding that think through every one of the risks that could possibly happen. And these are things like a robot tripping and falling and landing on your foot or pinch points, you know, where you might make contact or um I don't know, the arm swinging and and catching somebody right in the face or things like that. So, how do you then go through and mitigate all of those risks that you can imagine could happen? And I've already pointed out one or twice the physical barriers.

[00:11:14]This is the first step. But when you put the robots behind a physical barrier like this, you really eliminate a lot of their use. The whole point of a humanoid robot is to do many different more general purpose things. That's what AI, the promise of AI is more general purpose robots. If you have to put it in one work cell behind these barriers, you have to install something. It's no longer a robot just work walking and doing doing use. That's why they haven't scaled just yet.

[00:11:42]But what's coming very soon is a better approach where you have sensors on the robot that can reliably detect the people around. And by reliably I mean certifi certifiably like third parties certify that 99.99 something% of the time you can identify that a person is approaching and then the robot can change its behavior and stop moving and set down the thing it's carrying and sit down on the ground so that by by the time a person does touch the robot the robot is in a very safe state with its power off. This is a good way to make sure that we're not going to be able to injure anybody.

[00:12:21]What we really want though, of course, with robots being able to operate in our homes is this intelligent decisionmaking that we trust of how the robot is going to behave around people so that you can work right next to a robot so that you can shake hands with a robot so that it can hand something to you and you have very high reliability that that robot is not going to do something surprising or not going to be causing a problem. This is a big AI effort to do this. Think about how autonomous vehicles need to identify bicyclists and pedestrians and make decisions. That's a much more structured environment and it's only two degrees of freedom. You've got a steering wheel and the gas and the brake. That's it. A humanoid robot can do so much more in such more complex environments. So, this is a much harder problem to get to the point of the robots um making good choices. But we have to collect the data, right? We have to collect the data with steps one and two on the safety kind of pathway in order to build that trust in order to be able to show that these robots will have good judgment. So on that path, right, we start with fairly simple structured tasks like working in this warehouse and moving plastic bins around, which is what Digit does today as just the beach head market. But then there are hundreds of markets on the way. As they gradually get better at manipulation, get better at safety, they can start to work like in the back room in grocery stores and retail and then come on out and start stocking shelves, they can work in construction, hospitals, hotels, all different aspects of commercial and industrial life and eventually work their way in the home. And now they're kind of doing all of these things. So, think of this like a uh a foundational technology.

[00:13:59]And you know, instead of jumping straight to the end, maybe an analogy here would be uh electricity. Uh one of the most valuable things that electricity does now is computing and data centers.

[00:14:12]But but you start with a light bulb, right? You start with the first thing you can do that really provides a lot of value. And then the data centers come later, but you're now doing all of those things. And we still use light bulbs.

[00:14:25]That's what it's going to be like with uh embodied AI, with humanoid robots.

[00:14:28]they're going to start in these use cases and start to make it out into our lives. So to achieve this big dream of basic human needs are met of a tireless partner that works with us to really amplify our ambitions. Do they need to look like us? Do we need to build humanoids?

[00:14:49]Um not quite. There's no reason they need to look exactly like us, but we are finding they need to if they're operating in human spaces. As we go through and design something from the ground up, we find that two legs, upright body, and arms are a really good configuration. So, to help me explain that, I want to introduce our robot, Digit. Come on out, Digit.

[00:15:19]And remember the AI stack, right?

[00:15:24]There's DJ.

[00:15:31]Now on that AI stack, remember at the very top end we have the cognitive intelligence. In this case, this is a human in the loop. So I have a robot operator in the back because I don't trust Chet GPT yet to control this robot on stage with me. I do trust my uh my colleagues though. But all of the skills and the coordination um all those parts of the AI the physical AI are running on the robot right now. Um and uh obviously live here. Okay. So why do we need to have legs? Let's start there. Uh these robots are made to be humanentric which means they operate in human environments and we want them to do work. We want them to lift things off the ground, put them on the top shelf, and they need to have a small footprint in the in the ground to go through doorways and aisleways and things like that. That's what human environments look like. And that's a lot of the work that the robots need to do, which means they need to be balancing. If it was just a statically stable robot like a stool or something, and you lift up something kind of heavy, they get really tippy. But this robot is always balancing. Uh, and that turns out then that legs are the most stable way to be dynamically stable.

[00:16:42]Yeah.

[00:16:45]Hard to do that on wheels, right? Uh, but there's a lot of things that that are harder to do with wheels than legs, surprisingly.

[00:16:53]So, all right. So, given that we can be more stable on legs, uh, and more agile certainly on legs than we can on wheels, what what about the torso? Um, you remember we operating in the narrow spaces. If you're a velociraptor and you've got this long big space, when you turn around, you're going to knock everything over. Uh, so if Digit needs to turn kind of like in the aisleways, having that upright torso is also a really important thing. And there's a lot of stuff to fit in there, a lot of batteries and computers and everything else. So, it's got to go somewhere.

[00:17:24]Uh, and the other thing the torso is really good for is helping with balance along with the arms. You know, we swing our arms around when we're trying to balance. And when you start trying to walk, you lean in the direction you want to go. And Digit does the same thing.

[00:17:34]That's part of being dynamically stable is being able to balance.

[00:17:41]Now, the other thing that arms are good for is catching yourself if you're going to fall, which happens now and then for something that is dynamically stable.

[00:17:50]But these arms can reach anywhere around the entire robot, side to side, fork, front to back, whatever, to really make any fall as gentle as it it can be. And another question is maybe why two arms?

[00:18:02]You know, we could use an elephant trunk or something. Well, one manipulator at the end can only pick up like small things. And if you add another arm, now you can pick up much, much bigger things in addition to all the small things.

[00:18:15]Or you can hand one thing to, you know, one hand to another, reposition it, things like that. So having two arms and two hands is really important for a wide array of of big manipulation tasks.

[00:18:25]And finally, I would say it is a human-centric robot. This robot is designed to work with us and interact with us and on our terms, right? It goes into a human environment and does human workflows. Should also communicate with us well. And that includes having eyes and facial expressions and body language, right? And things that we can understand that mean something to us that is on our terms. We're not trying to program the robot.

[00:18:54]All right.

[00:18:56]So, at the start of this talk, I told you that this era of human centric robots is arriving now. I think you can see that it is. And I'm here to tell you, it's going to be great.

[00:19:09]Thank you.