Figure's Robotics Breakthrough for Humanoids

SPEAKER_1 (0:00)

I just finished recording this full podcast episode and realized that my microphone was not turned on. So I get to record it again, and maybe I'll be more practiced or I'll add a little bit of extra insights in this. Hopefully, it's even better if I talk really fast. I've already said this all once, so that's probably why. Today on the podcast, I want to talk about one of the biggest unlocks in humanoid robotics that we've seen to date, a massive technological breakthrough. It's honestly something that I'm very excited about, and this is all coming from Figur, one of the biggest humanoid robot companies that's competing against Tesla's humanoid robot. There's a handful of others, but Figur is one of the big leading ones right now. They've created basically a massive breakthrough that allows their robots to do generalist work. Basically, you take it and put it in a new environment and it's able to adapt now. So I want to break down how they're able to accomplish this. They built some really incredible solutions to problems that no one else, or I guess maybe other people had these problems.

But before we get into all of that, I wanted to mention that if you haven't already, I would love for you to go check out aibox.ai, which is my very own startup where you can get access to over 90 different AI models and you can work with them all in the same chat. If you're sick of paying subscriptions, $20 a month to 10 different platforms for audio, image, video generation, you can get all of the best models from Google, OpenAI, and Tropic all in one place. And if you use something like Claude for working something like Claude Cowork, you can actually get it to generate images right inside of Claude or video or audio with the AIbox connector. So it starts at $8.99 a month. We have higher tiers if you have really big projects, but I hope this is something that saves you a ton of time. I recently had a project where I got it to generate 1300 images right inside of Claude. It was able to do the whole project for me. I didn't have to mess around with it because it was using the AI Box Connector. And when it did that, I saw a massive spike in traffic to my website from that product and from that project that it worked on. So if you want to give it a try, there's a link in the description to AIBox.AI. All right, let's talk about what's going on with Figure. So yesterday their CEO put out a tweet where he said, hey, we have a massive basically unlock that we're going to be sharing soon. I was on pins and needles and today we got the drop about an hour ago. And by the way, if you're watching this on or if you're listening on Apple Podcasts, you can watch this on video. We have video on Apple Podcasts, and I'll be sharing some clips from their actual robot, what it's actually doing, et cetera. But for everyone else, I'll explain. So they released a new thing called Helix 2.5. This is their new AI model and a new update to their robots that let it do general purpose cleaning. Meaning in the past, the biggest problem that they had was that they could get their robot to clean a house in a familiar environment that had been trained on right. It's seen training data of this particular house and with particular objects that had already been trained on. So if you threw a bunch of objects on the floor and the training data had those objects in it, it knew how to pick up those objects off the floor. But if you had new objects, if you had new environments that it was unfamiliar with, it was unable to do this. So there's kind of this generalized issue where you could get it trained on specific environments, but it didn't go to all environments. They came up with a really creative solution, which I will explain in a second. But first, I want to tell you the results of the solution. The results of the solution were that they were able to take this and they did a big update. And to test this out, they rented 30 different AirBnBs across the Bay Area. So all homes that this robot, that the figure robot had never seen before, had never been trained on before, and they threw new objects in the home. So they would throw toys all over the floor. They would go mess up the beds, beds they had never seen. They would go put piles of laundry down. It was able to go clean the house, pick up all the toys, make all the beds, fold all the laundry in completely different environments where it had never seen these different objects before. So this is the generalized cleaning. They were able to do this on 30 different homes to validate that this model actually works. How are they able to get this result? This is what I was really excited about. So the way that these models get their training data, typically and traditionally, has been specifically for these robots. There's a couple different ways. One of them is through motion capture. So essentially that's a person wearing a special kind of suit. You use the same technology for a lot of animation, but it's got little dots on the suit, and you even have special gloves you can wear, and you go and do a task with a camera, watching how you do everything with all of the sensors so it understands exactly how you move. That can transfer to a robot really well. The problem is what I mentioned earlier, where it's just specific environments with specific objects that it's really good at.