00:13:39 · Brennand Pierce
[00:13:39] S9: Exactly. So this is why robotics is now doing so well. So we've actually been able to, as an industry, piggyback on the work that's done by Google, Meta, these big hyperscalers where they train these large language models and also with the vision backbone, which effectively, if you take a photo on your phone, it can tell you exactly what the objects are. With these two technologies, it actually unlocks the brain of the robot. So now the robot can finally know that a cup's there. You can say to it, where is the cup? And it can point to it. This was really missing. The control side, I. How to move the arm, how to control the robot, and how to balance the stuff, that is done still primarily with classical control that's been just like incremental improvement over the last twenty years. But it's really been this understanding of the world and getting a robot to actually know what it's doing. Before that, you would actually have, you know, 20 roboticists who would have to sit there for like three months and hard code every single line of code like do this, do that, do this, do that. Now with a mixture of using like ChatGPT or Claude, you can just tell it like this is what we want to do. Amazing. It can write half the code and the other half is what we call a foundational model. They're like these big hyperscalers like Google, Nvidia, they train these models by collecting hundreds of thousands of hours of a robot actually doing a task. So you have a human in the background teleoperating the robot. So they'll wear gloves or there's different ways of, like, teleoperating the robot. And someone will just, like, pick up a glass 10,000 times and then use this massive amount of data to then teach the robot how to pick up a glass. And the idea is if you pick up enough glasses in the world, the robot then knows how to pick up a glass.
00:15:23 · Katie Prescott, Danny Fortson, Brennand Pierce
[00:15:23] S4: Is it? No. That's really well explained. Thank you. [00:15:25] S8: And is this like... Is it effectively like the Tin Man? [00:15:28] S4: Very rude. [00:15:29] S8: But, like, it finally has a brain. Right? Like, the Tin Man can go out into the world and actually start to do stuff. And the reason I ask is because the Figure Robotics, speaking of very, very hyped humanoid robot companies, They announced something a few weeks ago where they're like, hey. Who wants to sign up to wear, like, a camera on their forehead and just record themselves doing tasks around their house so that we can use that data to train our robots? Was that what has been happening kind of... But on a very large scale throughout the industry? Like, is that... Is... Like, I'm just... Because it seems almost ridiculous, but I don't know if that's actually the way this has to develop. [00:16:07] S9: It's even more ridiculous than that. So for the last couple of years, you would literally have a room of three, 400 people holding grippers and manipulating objects and collecting. That's how you... We've collected. [00:16:19] S8: Oh, you mean... So you'd be like that, like, you'd be holding them as a person? [00:16:23] S9: No. You'd literally have, like, a pair of grippers in your hand, mechanical grippers, you just pick up objects. Or other ways people have been using it is literally like controllers. So you just control the robot like a puppet. [00:16:36] S4: Oh, right.
00:16:36 · Katie Prescott, Brennand Pierce
[00:16:36] S4: Oh, right. [00:16:36] S9: And you're just puppeteering the robot, and you might have, like, a 100. I don't know how many Google has, but they probably have, like, a thousand people just... [00:16:44] S4: Real people. [00:16:44] S9: Real people tying shoelaces, folding shirts, doing all these tasks, it's basically called imitation learning. It takes about one to 2,000 examples for the task to to basically learn any task. But then to make it really general, so going back to folding clothes, it probably takes one or 2,000 examples to fold your normal t shirt. But if you want to fold or do any laundry in any houses, no one really knows. Like, you might need a hundred thousand hours worth for someone doing laundry, then you could generalise it. [00:17:19] S4: We have covered this in a in a different capacity before when we looked at WAVE driverless cars, because they got footage from Ocado vans and Astavans of of people driving, essentially. And and that's how they created their model. So it's quite similar.
00:17:19 · Katie Prescott, Brennand Pierce
[00:17:19] S4: We have covered this in a in a different capacity before when we looked at WAVE driverless cars, because they got footage from Ocado vans and Astavans of of people driving, essentially. And and that's how they created their model. So it's quite similar. [00:17:34] S9: It's exactly the same. The advantage, like, large language models had was they could scrape scrape the Internet. Another way that, you know, the AI used to be trained, if you used to go to websites, and it would literally ask you what is a bucket, and you had to click it. Yeah. And that was you training an AI model. So that's how they could train the image classification, how they could learn language. For a robot, you basically have to get a robot to do the task. Historically, that is it. As you were saying before, Figr and some other companies claim that you could just wear a camera on your head. I don't know if that's been proven, but I think it's like all this mixture. You still need a lot of different data coming in because all the things you need is like the contact forces of the... A griffle, like how how much do you squash something if you're picking up Hold an egg. Holding an egg. [00:18:19] S4: Or a grape. [00:18:20] S9: A grape. Or you're picking up like sushi compared to like, I need to pick up a bolt. So you need all of this, what we call multimodalities. So you need forces. You need cameras. You need sound. You need all of these sensors to come into the robot, and you just need an incredible amount of data.
00:18:20 · Brennand Pierce, Danny Fortson, Katie Prescott, S0, S1, S2, S3, S5, S6
[00:18:20] S9: A grape. Or you're picking up like sushi compared to like, I need to pick up a bolt. So you need all of this, what we call multimodalities. So you need forces. You need cameras. You need sound. You need all of these sensors to come into the robot, and you just need an incredible amount of data. [00:18:42] S8: Today's episode of the Times Tech Podcast is sponsored by ServiceNow. [00:18:46] S4: ServiceNow describes itself as an AI control tower for business reinvention, connecting people, systems, data, and AI so work flows more smoothly across an organization. [00:18:57] S8: The platform integrates with different clouds, models, and data sources, bringing them together in one place to help your team. [00:19:03] S4: To learn more about how ServiceNow puts AI to work for people, visit servicenow.com. [00:19:12] S0: People are disappearing again just like before. [00:19:17] S1: Nocturne. Streaming on Apple TV on the October 30. Lief Schreiber. [00:19:23] S2: Walter has a way of getting people. When he does, [00:19:25] S1: they don't come back. Steven Graham. The world [00:19:28] S3: is full of dark places. Zarcee Beats. [00:19:31] S4: How's he getting messages at? [00:19:32] S5: He's locked under six feet of concrete. [00:19:34] S0: She's still alive. [00:19:35] S6: I'm a fighter. [00:19:36] S1: Nocturne. Streaming on Apple TV on the October 30. Subscription required.
00:19:36 · S1, S7, Katie Prescott, Brennand Pierce
[00:19:36] S1: Nocturne. Streaming on Apple TV on the October 30. Subscription required. [00:19:42] S7: This episode is brought to you by Capital One. Capital One's tech team isn't just talking about multi agentic AI. They already deployed one. It's called Chat Concierge, and it's simplifying car shopping. Using self reflection and layered reasoning with live API checks, it doesn't just help buyers find a car they love. It helps schedule a test drive, get preapproved for financing, and estimate trade in value. Advanced, intuitive, and deployed. That's how they stack. That's technology at Capital One. [00:20:18] S4: Brent, we will get some pictures of your robots for people watching this on YouTube. But for people listening, describe your robot and also explain to us how you came up with that design. Because something I've noticed about the humanoid robot world, which you two will be very familiar with, is they don't always look identical. They don't look like, you know, one might expect. [00:20:39] S9: Yeah. Even though my passion is like humanoids with legs, one of the things I learned over the years is money matters when it comes to selling these. So having like, I think it's about 14 or 15 expensive motors to actually make a robot walk around is cool. But in reality, if you just work in a warehouse or factory, it is much easier just to have it on a wheeled base. So it has like two arms. [00:21:04] S4: I completely agree. [00:21:05] S9: And the torso just goes up and up and down. [00:21:07] S4: I've long thought the leg thing. It's just completely ludicrous.