Science Friday
Hosted by Flora Lichtman and Ira Flatow
Roboticist Karen Liu combines machine learning and animation to teach robots to move and respond more like humans.
3,146 words · 8 speakers
Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
Science Friday is supported by MathWorks, creator of MATLAB and Simulink software for technical computing and model-based design. MathWorks, accelerating the pace of discovery in engineering and science. Learn more at mathworks.com.
Hi, this is Flora Lixman, and you're listening to Science Friday.
Today in the podcast, a conversation from our live show in Redwood City, California. Ira's talking AI and what it means for humanoid robots.
The conversation part, it feels very amazing, but it's actually the easiest part.
You know, robots are just about everywhere these days. Maybe you've spotted one circling the grocery store, cleaning the floor at the airport, making deliveries, not to mention the robots on the assembly lines, factories. But what about having your own robot in your own home? How far are we from having a human-like robot at home? Joining me now to talk about the latest advances in humanoid robots is my guest, Dr. Karen Liu, professor of computer science at Stanford University, based, of course, in Stanford, California. Welcome to Science Friday.
Thank you, Ira.
You're welcome.
Nice to be here today.
Now, you developed and trained humanoid robots. What is the definition of a humanoid? How do you define one?
Well, there's really no universal consensus on the definition of a humanoid. But I think most people would agree that, you know, humanoid should have a similar morphology to human. That means you should have two legs for locomotion, two hands for manipulations, and you need to be able to have egocentric vision systems so you can navigate the world, you could interact with the objects in the world.
Now, I understand not just talking about it, you've actually brought one.
Yeah.
Recently, right?
We actually built a humanoid from scratch at Stanford.
Toddy.
Toddy. As we call it toddler bot because it's model off of a human toddler. It's very small, but it has a lot of features, almost like a full-size robot. It can talk, it can listen, better than toddlers, and it can also see the world.
All right, let's bring it about. Let's bring out Toddy the toddler robot. Toddy, are you excited to be here tonight?
Hehe, I'm so happy to be here too, buddy.
Toddy, can you tell us how old you are?
I'm one and a half years old. Yummy.
Toddy, do you know that you are a robot?
Hehe, yep, I'm a robot toddler. Wiggle wiggly.
I have to ask you something I ask all my two-year-old grandchildren. Toddy, do you have to go potty?
Hehe, nope, no potty for me. I'm all goodie.
Thank you for coming, Toddy. Good to see you. Thank you. Nice to see you, Toddy.
Aw, nice to see you too, buddy.
Thank you. Isn't that great? That's terrific. Tell us, what are some of the advances in robotics that make something so humanoid like that?
Yeah, so there are so many things. You know, AI is definitely one thing. I mean, probably heard of that, you know, ChatGPT, large language models. They fundamentally change the way robots think about the world, reason about the world. But there's also something that is, you know, less mentioned, you know, often overlooked factor, which is the advances in hardware, in robotics hardware. And this is not so much about like one single breakthrough, technology breakthrough, but rather a convergence of all the enabling technologies. So for example, 3D printing, like the toddler body you just saw, they are completely made of 3D printed materials and all the motors are commercially available.
So this is something that you can actually just build from scratch at home. And then I think that is one of the reason that you will start to see that that this is very welcoming environment for robotics hobbyists to do this.
How do you how do you teach it to be intelligent like that?
Yeah, so that's a really good question. The conversation part is it feels very amazing, but it's actually the easiest part.
Really?
Because of the chat GPT, because of large language model. If we have an interface that can communicate with chat GPT with a large model fast enough, then you can, you know, basically directly talk to chat GPT. And so what you just did, we didn't do any scripting, right? I have no idea what IRS going to ask.
Right, you didn't know I was gonna ask.
I was a little worried that.
I was more worried than you were.
Yeah, I was. But I think the hardest part is the something that you feel like very easy to do. For example, balance, for example, walking forward, locomotion, manipulation. Those are things that we we take for granted, but it's very difficult to train robots to do that reliably.
Is that the hardest part then, to get it to walk and and and to mimic how humans do things?
Yeah, so walking is hard. Manipulation is another very difficult task. And the reason walking is really challenging for bipedals, meaning the robots with two legs, is because we are, this is what we call underactuated system. Underactuation means that, you know, we could actuate our our joints, our motor, well, we don't have motors, but human noise use motors. We could actuate the motors locally, but the whole system, the whole dynamic system, we cannot accelerate it from point A to point B without using external forces. So the way we taught of human noise can walk is because it figures out how to exert torques
at its hip joints and knee joints and ankle joint in a very coordinated way so that you can push the ground, get exactly the forces you want from the ground and move your accelerate your center of mass forward to the right location that you want it to be. And that's a lot of a computation there.
I'll bet. What about, I imagine the hands are just as difficult, right?
Yeah, hands, that's a that's a different thing. And the reason manipulation is really hard is because we really want to be able to manipulate. Anything in the world in a very general ways, right? If you just build, you know, robot hands, robot arms to do one thing in the in the factory, in the industrial setting, that's actually not too hard. But the reason it is hard today is because manipulation means that you need to handle all different kind of diversity in the world and do so with the same brand, same policy. That's what we call it.
Well, what do you think the limitations are of what you can teach Tati or robots to do?
One of the biggest limitation is that robot doesn't, you know, like if you were going for the machine learning or AI approach, we need a lot of the data to teach robots, just like the way we teach, you know, large language models. So my colleagues would say, we are, we have this 100,000 years of data deficiency comparing to training large language models. What it means is that, you know, in order to train a LLM or large language models, it takes the data that that takes an individual, a person to read the text, training data, 100,000 years to read.
Wow.
And if you think about all the largest data set we have in robotics, it's about 10,000 hours.
10,000 hours?
Yeah, that's the data, amount of data we have today.
Right.
So, so data is definitely really challenging for robotics. And another, you know, really difficult things just we're learning, not just understanding text or images, we need to understanding the mapping between what you see and what you do, right? Because this is the most crucial part of robotics, a decision maker.
Now, we have lots of people lining up. I'll try to hit both sides of the room. Let me start on this side first, yes.
So you've talked about data to teach the robot. I'm a science teacher who teaches life science, and I think a lot about how our senses give us feedback, which is like a huge part of learning how to walk, for instance. What kinds of feedback does the robot take?
Yeah, so using data to teach robots, we call it imitation learning or supervised learning, is only one way. Robot can also learn from trial and errors, learn from its own experiences of interacting with the world, and this is what we call reinforcement learning. In that case, we will ask robot just to try things, you know, here's your task, you try different actions, and then we look at the results, we'll give you a score to let you know you're doing well or you're not doing well, and robot will take that as a signal to decide, oh, what I just did was pretty bad, I got a four score, maybe I should reduce the probability of doing the same thing again. So that's kind of like, you know, getting feedback from, from human.
But in order to make this process automatic, we don't have a human, you know, keep telling robots, you know, what score you get. Instead, we, we will design what we call a reward function. It's an automatic way to assess the, the, the performance of robot.
Nice question. Over here on this side.
So to follow up on the concept of the reinforcement learning, I read as a young boy Isaac Asimov's books, and among them were Rules of Robots. And so how does this reinforcement learning coordinate with that and the ethical use of robots to be sure they're not going to harm us?
Yeah, so, you know, talking about ethical aspect of robotics, it is still, so the reward function I was just talking about is designed by a person, right? So, you know, if we believe there are certain values and believe that we want, we want robot to learn from, we have to encode it into the reward function. And that is definitely not an easy way to do because, you know, how do you turn ethics into a mathematic function in the equation?
Do you think robots are going to get smarter than us and we'll be the robots someday?
Well, you know, robots are, can appear very smart in certain ways, but sometimes they were also not so smart. So for example, going back to that reward function again, my colleague was, you know, would say, you know, if you ask a robot to cook dinner for you and you write it down into the reward function, robot would probably do that, cook dinner for you, but along the way, it kills your cat. And the robot say, well, in the reward function, you say, you know, I got a good reward for cooking the dinner and I don't get any punishment for killing the cat, you know, why not, right? So the problem is that you can never, it's kind of like arguing with a toddler. Like you didn't say I could, it cannot do that, right? There always the, you know, corner cases that you cannot capture in your reward function.
We don't want to kill any cats, no. What happens when a robot fails? It can fail in many different ways, right? Can it actually teach itself what it did wrong without you having to tell it? Does it learn from its mistakes?
Yeah, that is actually a really great question, Aya, because I would say this is the major deficiency of a robot today comparing to humans. Humans are really, really good at knowing how well, like self-awareness, like knowing how well it does in this particular task during the skill learning time and after skill is deployed to the real world. If you ask a small child, you know, to put together a block structure and, you know, if the child fails at the task, it immediately knows what went wrong. It knows, oh, maybe I didn't set a foundation sturdy enough
or I didn't align this particular structure, piece well, precise enough or a dog just nudged a table and, you know, caused the structure to collapse. They know immediately. And this ability, this innate ability of understanding the source of error is something that a robot doesn't have. So a robot will fail it. And then the only thing it would know, it has no idea why it fails. It just, well, assume, I'm not going to do that same thing again because I got a bad score. But if a robot can reflect, reflect on its behavior, maybe it will figure out, oh, because my sensor was not calibrated right, or I estimated the floor not as slippery as it should be.
And then knowing that will help it to accelerate learning a lot.
It's not going to say, oh darn, I did something. It's not going to go that way.
After the break, more of Iris' conversation with roboticist Karen Leo about the current state of humanoid robots.
Yeah, so if you want something like Tati, you can have it today.
You can.
Oh, yeah.
WNYC Studios is supported by Wise. Wise, the smart way to manage the currencies you need around the world. With a Wise account, you can send, spend, and receive in up to 40 currencies with only a few simple taps. Be smart, get wise. Download the Wise app today. Terms and conditions apply. Science Friday is supported by PBS, presenting Going Wild. If you love learning about the natural world and meeting the scientists who help us understand it, be sure to check out the new season of Going Wild with Dr. Ray Wynn Grant. It's the chart-topping and Webby Award-winning podcast from the folks at Nature on PBS, now with a video companion series. In this season of Going Wild, host acclaimed wildlife ecologist Dr. Ray Wynn Grant
heads out into the field to explore the animals, people, and stories that make nature worth saving. Stops include a hibernating black bear den in Arkansas and an underwater shipwreck site off the Florida coast. If you're looking for a place to start, check out the season's debut episode, where Ray joins beloved biologist and influencer Jordan King in the Florida Everglades to learn about his unlikely journey into science and how to study a wetland when it's not very, well, wet. So don't miss out. Follow Going Wild with Dr. Ray Wynn Grant on Apple Podcasts, Spotify, YouTube, or wherever you get your podcasts.
And from the Alfred P. Sloan Foundation, working to enhance public understanding of science, technology, and economics in the modern world.
You have a background in computer animation, correctly?
Yes.
How has that helped you in your robotics? What did you bring to that?
That's right. So I actually started as a computer graphics person and then kind of worked my way into robotics. But if you think about computer animation, you probably imagine there is an animator making those key frames, those poses, and then they interpolated those key frames and make a nice-looking, smooth animation. And that is something that you can do for, as I say, making movies. But if you want to do video games, then you probably will need a process, an automatic process, that can generate motion. Let's just call it motion generator, which understand the high-level commands like moving forward, make a jump on the platform, you know, when you play video games.
And if you take one step further, if this, you know, instead of this motion generator, instead of outputting a pose or a sequence of poses, it outputs joint torques. Then what you can do is you take that joint torque, together with the gravity and contact forces, whatever the external forces is, put it into a physics engine and let the physics engine figure out the poses. And this is what we call physics-based animation. And at this point, you are actually very close to robotics. The only difference is that you have a human animator that giving those high-level commands, or you have AI models to do that.
So if you replace the humans with an AI model and let this AI model make high-level commands based on the task, based on the observation, then this is basically, your character is basically a virtual robot, and you are ready to deploy it to the real world.
Now, I understand that Tati, who was at before, was actually 3D-printed?
Body?
Which makes me think, how far along until we can actually get our own home robot, do you think?
You can?
Do we have them in the lobby for sale?
No, but almost as good as that. We have, we open source everything. We open source the hardware design, the algorithms, and also step-by-step instruction manuals. So if you have a 3D printer at home, you want to build Tati, all you need to do is just order those commercially available motors and then just follow the instructions.
Sure. If you can bake a cake, you can do that, right?
We have the recipe.
But in general, when might we see home robots? Not maybe just Tati, but other robots?
It is really difficult question, and I get this question all the time, right? And I always say five years.
Everything is always five years away, right?
That's right, because beyond five years, then like, you know, it's not my responsibility. The reason I think it is possible, but not just, you know, tomorrow, is because right now, we're still building a lot of building blocks, really, like locomotion manipulation. Before, they are completely reliable, robust, and safe, I don't think it is ready to build a product out of it. But I also feel like the deployment of humanoid is going to come in stages, kind of like, you know, autonomous vehicles. So at very beginning, maybe the robots are not going to be able to do anything you want
in a household, maybe just for laundries, or, you know, cleaning your dishwasher.
I'd settle for that.
I know. Those are the two things I want.
Dr. Leo, this is fascinating. Thank you for taking time to be with us today.
Yeah, of course.
Dr. Karen Leo, professor of computer science at Stanford.
Thank you so much.
Face of course in Stanford, California.
Thanks for listening. Don't forget to rate and review us wherever you listen. It really does help us get the word out and get the show in front of new listeners. Today's episode was produced by Shoshana Buxbaum. I'm Flora Lichtman. Thanks for listening.
Page updated November 1, 2026 · transcript and summary are machine-generated