Figure’s Humanoid Factory Tour – CEO Brett Adcock

[00:00] Welcome to Figure. We make humanoid robots here. We design them, we build them, we test them all here. This is a secret room that nobody's allowed to come in. I'm going to show you every robot we've ever built. So here we lost like a left knee and you can see the robot's kind of like hobbling on the left leg. So we have a robot doing burpees here. This is where we manufacture Figure 3 robots. Wow. And this was the first car in the world built by a humanoid robot that we're aware of. [00:30] autonomously from an onboard AI policy called Helix. Give it a push. Oh gosh. Okay. I feel bad. [00:48] Welcome. Hi. How you doing? Good, how are you? Good to see you. [00:52] - I'm excited. - Yeah. [00:54] Welcome to Figure. [00:56] - Thanks. So where are we right now? What is this? This is the headquarters? - This is a robot campus. [01:01] - Robot campus. - Robot campus. - We make robots here. - Humanoid robots. - We make humanoid robots here. [01:05] We design them, we build them, we test them. [01:08] - I'll hear. - So for people who don't know what a humanoid robot, you will see in a second, 'cause this is kind of freaky, [01:15] Can you explain what they are? Yeah. Our goal is to build advanced AI that we can put into a general purpose [01:23] a humanoid body. [01:24] A humanoid is basically just like a robot with a human form. [01:27] So we have arms, hands, head,

[01:30] feet, legs, we can basically do everything a human can in the world with one piece of hardware. So, yeah, our goal is to be able to go out and [01:39] basically design and ship humanoid robots in the world that can do everything from housework, dishes, laundry, to manufacturing, healthcare, just basically as much things in the world as possible that we can go out and ship robots too. [01:53] And we see this on the screen. [01:55] Oh yeah, that's our-- [01:57] - Is this your hype machine screen? - Yeah, exactly. That's our latest generation robot figure three. [02:04] Do a little like a concierge job. [02:07] Cool. [02:07] Okay, so what are we going to see today? What's the plan? Okay, come on in. I want to show you some of my robots. Okay. I think first here is we have some robots that we basically have constantly running around the office 24/7, talking to humans, greeting people, just basically doing useful work. [02:27] These robots basically can run fully autonomously without any humans, and they can automatically dock themselves and charge. [02:35] And then once they're fully charged, they basically come off and be able to do useful work. [02:40] So these robots that are docking here, [02:41] are charging through the feet. [02:43] So we have a wireless charging stand, like similar how you would charge a phone, like inductively. Okay. The robots can charge through their feet at two kilowatts. [02:51] So basically, the battery lasts about four to five hours, then we can charge basically for an hour and go back and do work again. [02:59] So the robots, we don't need to do anything. We don't need to plug them in.

[03:01] They can basically auto charge themselves and just do 24/7 operations. Is the role for this one right here just for the docking exercise? Do these ones roam around? These ones roam around. Yeah, they're just docking here just to charge and then they'll be out doing useful work all day. [03:17] Bye. [03:18] And I saw these ones as well. [03:20] Is one of these special? Yeah. So this robot right here, the one with the American flag, was actually at the White House last week. How did that happen? We got a call asking to be basically the first humanoid group, basically [03:36] ever to put robots in the White House. [03:39] And so last week we basically had the first robots in-- humanoid robots in history there. [03:45] I THINK IT'S A LOT OF PEOPLE. [03:46] doing stuff, talking, greeting folks. We were basically at a very special event with the First Lady. [03:52] And it went really well. [03:54] It's exciting. [03:56] So this is our corporate headquarters. We have four buildings on campus. [04:01] So here we do a lot of basically engineering design work. [04:05] How many people work out here? We have about 500, a little over 500 people. How many are in the total company? Um, five, oh, sorry. There's about 250, 300 people here, and we have 500 in the company. Oh, wow. [04:17] Yeah, I would say most all of its engineering, and then we've been growing out manufacturing. [04:22] supply chains, some of the areas to basically, how do we make more robots pretty aggressively. And how many robots do you have here? How many ones are there? A few hundred right now. Okay, are they outpacing the humans or? There's so, my goal for the, for this building is I want more robots, humanoid robots than humans. Okay. Walking around and they don't necessarily need to walk. They could be sitting or talking. Mm-hmm.

[04:43] Thank you. [04:44] Yeah. So basically we have, we basically do a lot of, [04:49] basically hardware and software validation testing here. [04:52] Okay. So like testing for burn in. [04:56] durability, [04:59] Basically, like any new software or hardware gets validated through this facility. [05:05] So you have robots doing all kinds of crazy stuff here. [05:08] That is definitely some yoga. [05:11] Robot getting on the ground and getting back up again. [05:14] We're basically trying to stress test the robots to try to find any potential failures. Okay. For like any new hardware software could get released. So if we have a new camera, [05:25] a new type of like, say, structure or anything else, we'll test it here before it goes out. How many potential body movements can they do? Okay, so this is kind of crazy. So the robot's basically made up of about 40 motors. Okay. Every motor can spin like 360 degrees, like all the way around. So, mathematically, how many states it could be in, like body positions, is 360 to the power of 40. [05:50] - What? - Yeah. It's more body positions than atoms in the universe. [05:55] Which is crazy. Potentially. Yeah, I've done the math. It's for sure. You've done the math? Yeah, it's for sure. Okay. So we, so it's basically like the difficulty here is like, how do you control it? You can't write code to make this work. [06:07] So all of our robots here run on a neural network we call Helix. [06:10] It's a vision language action model we designed here internally.

[06:14] to tell the robot what to go do, to stay balanced, like how to move its joints, basically, from pixels, from camera space. And so that team that works on Helix is in the same building? They're in the same building, right here. [06:25] Yeah, they're phenomenal. We basically have a large scale data collection efforts that's going on, and then we train our own models here internally. [06:32] And then we test them all here as well. [06:34] It's not just like, [06:36] It's not just for balancing and being able to have stability, which we need to have human-like stability. [06:42] It's for how do we know what to go do? [06:44] How do we take in prompts from humans and say, like vision from the cameras? And how do we output every single joint, including where the body's positioned and feet and hands to do stuff? Like it could be folding laundry, doing manufacturing, the logistics that you'll see here later today. And we have to do that like a few hundred times a second. [07:01] from camera images. [07:03] And on a neural network, everyone's on board the robot. [07:06] So it's a really hard podge order. Wow. [07:09] So yeah, all the bays here are running some sort of test. [07:13] for durability or reliability testing? [07:17] Um... [07:18] Why do some have different suits? [07:20] Oh, we outfit them. All the outfits are fabric. [07:26] Like a human, like clothes. Yeah. And they all have different clothes. [07:29] Which is cool, because we can, like, you can accessorize the robots how you want it. [07:35] our clients can have different outfits that show the client logos and colors. This is a new level of merch. Yeah, it's a new level of merch. Okay. It's also nice because if things get dirty or they rip or whatever, we can basically easily replace it.

[07:47] Without a technician. Yeah. Yeah. All the robots have a little zipper on the back. I'm sure you see. [07:52] We can basically just unzip it and basically take it off and put something new on. [07:55] - Okay. - Yeah. Also, it's just really cool. [07:58] - It is pretty cool. Their shoes look like real sneakers. - They're high tops. [08:02] - They're high tops. - They're high top sneakers. And yeah, they look, don't they look awesome? [08:07] Yeah, I mean, they look like [08:10] human bodies. - Getting the lab to this level of like, [08:14] infrastructure to be able to run them like this every single day is actually quite difficult. Then we need to be very diligent about when we find issues, like how to track them, how to do fault analysis really quickly, and then how to solve them and then how to solve them across the global fleet, basically our whole fleet. [08:27] wherever they're at. [08:28] How many are you... [08:31] in development testing all at the same time? Like what is the typical, is it, these bays are always active? How does it work? - Yeah, we basically, so the goal of this lab is to basically do final, [08:43] Final checks for all software [08:46] That could be embedded software, it could be a neural network, a helix. It could be firmware on the robot and then any new hardware changes we have. We need to make sure those changes are bulletproof before they head out of here. Because it's going to cause a lot of problems if like, [08:59] we're trying to run a use case for logistics or home. [09:02] The robots are messing up. We're not sure why it's messing up. That's not great for us. [09:05] So basically here, we're basically doing a ton of testing. We have like test plans laid out every single morning. [09:10] We're running those down and we see any potential falls, we have to go solve it. [09:13] Then we have to retest those plans. [09:14] So these robots run in here like all day.

[09:17] every single week. [09:19] And we run them really hard. [09:20] And the goal is we don't want to be finding failures upstream out of this lab. [09:25] Yeah, so this is the banner on here on system integration tests is [09:28] Trust but verify. Okay. So, um, [09:31] Yeah, we basically have to make sure we run down [09:34] every potential thing that could go wrong. [09:37] before leaving here. [09:38] This is like a-- [09:39] This is a hard thing. The robot has 40 plus moving joints. It's like a walking cell phone, self-driving car. [09:46] All the supply chain is basically new. We've designed almost all of it. [09:49] And so it's hard. It's hard to get the system to be really reliable. [09:54] And it's not like if we lose power, [09:56] We're not like, like us, like, [09:58] We're not statically stable, so if we lose power, the robot falls. [10:02] So we can never lose power, we can never lose comms. [10:05] and then we need to be able to balance at all times everywhere we're going. Even if we're moving the body, like your pelvis and hips and everything are moving in relation to your, like, you know, as it relates to your hand moving and things like this. [10:14] and head. [10:14] So it's quite a difficult problem. [10:17] And so that's why you dock them at 15 minutes. [10:20] Or 15%. [10:21] We dock it around 10 to 15%, they'll go to dock. [10:24] And then if we need another robot in, we'll sub him in off the dock. [10:27] And you'll see here later in our logistics and other use cases that need like constant, like 24/7 attention, the robot will undock. [10:33] right before the other robot needs to leave. And the robot will then basically do a quick swap [10:38] and within like 30 seconds it's now doing work, another robot will go in and dock. [10:42] And we'll just run that every four or five hours on repeat 24/7. [10:46] What's [10:47] - What's the most common error that they make? - Most is software at this point. - Software? - Yeah. The hardware's gotten really robust. I mean, the hardware here is great. We just, basically, it's a software issue.

[10:58] - Mm, and then, [11:00] In terms of the hardware, you're manufacturing also on campus? Yeah, we manufacture here at Baku next door. OK. Yeah, we're going to show you that today. OK. Yeah. OK, so we have basically a robot doing burpees. [11:12] Here. We want to be able to safely for any event, get down to the ground. [11:17] And then we want to be able to safely get up. Okay. It's like important case. We're like not sure what to do We could be on like a very low battery [11:23] and we might need to safely get down. We could have like, and then we want to be able to get off the ground really easily. [11:34] It's also quite hard. - Yeah. - You need a ton of like, like range of motion in the legs and the hips. [11:40] to be able to do this kind of maneuver. - Yeah, he's gonna need some knee pads. Is there a reason why the joints are hard and you don't have soft tissue? - Yeah, most of the upper body torso is all soft. - Okay. - And then we have some soft foam underneath the legs and arms right now. [11:56] Yeah. [11:57] Obviously, like the more... [11:59] The more patti and more soft I think is great. It just adds a different level. It adds more volume and mass to the robot. [12:05] Yeah. [12:06] It makes the robot look bigger, basically. [12:08] some thick robots. Yeah. [12:11] *gasp* [12:12] Cool. Okay. Let's go. [12:15] So how often [12:16] Do you come into the office every day and you check in on them? Like how do you... Every day? Do they feel like you're babies? Like do you feel like... They're for sure babies. Do you have a parasocial relationship with them? They're like, we've like made these, like they're little kids. And we have to like get them to do useful things now.

[12:30] I think the good news is we're at a point where the hardware's gotten almost very robust. Yeah. We can run them all day, every day. We still see hardware failures, but [12:42] It's very few and far between. Most of our problems now are like, as we think about this baby growing up, are software problems or AI problems. [12:50] How do we get the software incredibly stable and how do we get the neural networks to be able to actually do useful things 24/7? [12:55] without failures. Most of our failures today are [12:58] kind of in software land. [13:00] And speaking of software, [13:01] Moritz, one of our favorite things to do is push robots around. Okay. And so Moritz here is one of our leads on basically Helix controls team. [13:11] and is gonna, maybe you can give her a quick 101 of kind of the S0 controller we have here, and then [13:17] We would love for you to push the robot as hard as you can. [13:20] Yeah, so I think what we did recently switched fully to RL from a monolith stack, the branch, and there's this, that [13:27] that we have all this variation that we can give to the robot when you train it in SYN. So all edge cases are now known to the controller and gets robust to it. So [13:37] What this means for example, [13:39] Before a model-based robot stack got it freaked out, you have this very robust external services. We push it around. [13:48] Go ahead to convince yourself. Really, I think this really nicely showcases how robust our stack is. Give it a push. Oh gosh. [13:56] Okay, I feel bad. [13:59] That's-- - A little harder.

[14:04] Okay, so... [14:06] They don't know, what do they do if they get attacked? You know how Waymos get attacked? People were attacking bird scooters. [14:14] Do they have a defense mechanism? No defense mechanism. None? No. It's not trained in film? No. They don't see the internet and see what happens? No harm to humans. No harm to humans. No? [14:23] No, they're here just to help. [14:25] - Okay. - Yeah. They're here to take a push too if you need to. [14:30] - Yeah. - You wanna get another one in? [14:33] Sure. Yeah. [14:35] - [14:36] It feels really heavy. - Yeah, yeah, yeah. It's like 135 pounds. [14:41] But it's actually really, it's got human level stability. [14:46] AND, UM, [14:47] As Moritz mentioned, we're learning that coverage in a simulator. [14:52] So the whole controller learns [14:54] how to stay stable like this synthetically, like in a sim. Basically like a video game. [15:00] And from there, the robot learns how to stay balanced. [15:03] how to basically not fall whenever there's certain forces, [15:08] and we basically can zero shot it onto this robot, meaning we can just put it right on [15:12] load it to the computer and we can basically [15:15] and get this level of performance in the controller. [15:17] What are like... [15:18] Again, what are the most common tweaks within that in the software? You're obviously balancing a lot of physical issues there, so how do you tweak that in the models? [15:30] Moritz, how do we get the models to be able to perform like this?

[15:34] Um... [15:36] Basically, I think we spend a lot of time thinking what are all the things that can happen to the robot in the real world and then make them happen in simulation. [15:44] I think that's the... [15:45] You basically have like, it's like a physics simulator. So it has like gravity, [15:48] has like friction coefficients. [15:50] We want to try to mirror like what forces robots seen here. [15:53] so we can run them in SIM. [15:54] And if they can run in sim well, what we've seen is we basically have a really great sim to real transfer. So we can get it from a simulator that shows like, you'll see it like in a simulator like video game. It can like, it can like stay stable with these forces. [16:05] Then we load it to the robot. [16:07] and we see the same in the real world, and we have like a basically a very high transfer rate. [16:12] - It's interesting, I feel like, [16:14] I feel like [16:15] It would be interesting to hear your perspective or differentiation on your robot versus the other humanoid robots and like, [16:23] where this physicality and [16:26] essentially the behavioral mechanisms [16:29] change is some might be more commercially focused. These ones are clearly [16:33] as humanoid as I've seen. Yeah, we're a pure play humanoid. OK. Yeah, no, I think a few things. [16:40] One is we need to get the hardware in a really good spot that can do a lot of what humans can do. You really want this iPhone moment where [16:47] My iPhone has a bunch of different apps. And if I wanted to learn something new, I just download another app. [16:52] - What is the same thing for humanoids? What is the same humanoid to be able to do dishes and laundry, but also do some package logistics and healthcare and other stuff? That's what humans can do, right? We're all fairly general purpose. [17:02] And so we want one set of hardware that we can amortize over a lot of different use cases.

[17:06] So the goal is to get the hardware robust enough to do most things a human can. In terms of range of motion, go get down, get off the ground, be able to reach up high, be able to reach inside a sink, all these different things we need to do. We also need to carry a decent amount of payloads and we need to operate at decently fast speeds. [17:21] So we've designed the hardware to be able to do that. [17:23] And then separately on the software, on the AI side, you really want to be able to design the neural nets so that it can basically take a task, [17:30] and then reason through pixel space. I take camera videos of what's happening, and then output what the body should be doing. [17:36] The [17:38] You know, the math we were doing before on like how many states the robot could be in is just, it's so high. Yeah. That like the problem just kind of runs away. It's, [17:48] It's like a curse of dimitiality. It's just too hard. [17:51] So you can't solve it with like writing like lines of code. [17:53] Like attritionally, like in robotics, [17:55] And even like three or four years ago here at Figure, like we would solve the same controller here in code. [18:00] we'd have like hundreds of thousands of lines of C++ to figure out how to solve this like inverted pendulum math of how to like stay balanced here. And what we found is it just doesn't scale. [18:11] It doesn't work. You can't really, in your head as humans, [18:14] work across many different humans that code all this stuff into the robot that you think it could encounter in the world. It's just too difficult. [18:19] We transitioned only purely to neural networks with Helix, our neural network model. For those, I know you mentioned this a couple of times, but this will be a general audience. For those that don't know what a neural network is, can you just explain that a little bit further? Yeah. We basically use an AI policy that we've trained here with data.

[18:37] We trained like a transformer transformer policy to basically output [18:40] a certain type of action space. Basically, we train an AI policy to do this work of what Koda used to do. [18:46] and learn this. [18:47] So now basically we run inference on board the robot. [18:51] across a policy that we call Helix. It's an AI model that we designed here internally. And that policy is outputting what the robot should go do. [18:59] a similar how you talk to like an LLM, and you'll ask it what to go do, [19:03] It'll do inference and output, like basically like a next token prediction for words. [19:08] We do the same thing here, but for a physical humanoid. And we'll output things like where to put the wrists, head, [19:16] torso, every joint will get an output from the neural network, like what to go do. [19:20] And we'll do that anywhere between 50 and 200 milliseconds. [19:23] So basically 50 to 200 times a second, the neural network computer [19:27] is then telling all the joints what to go do. Then every joint level, our motor controllers are outputting torque. [19:33] on where to send basically, like where to position the motors. [19:36] - Okay, wow. - So yeah, so basically, you can have two pass here, you basically can code your way out of this, and I think that's a full dead end. [19:45] or you can run an AI-first strategy in the market. And that's what we do here at Figure. [19:49] So I think what differentiates us is from a hardware perspective, I think we probably have, I think this is probably the best humanoid hardware in the world. [19:55] to do general purpose work. [19:57] And secondly, all the work that we're doing is all neural network based or AI based. We don't code any of this work anymore. [20:03] You'll see some use cases today that we do both for the home and for cases of a commercial market. Those are all run by our Helix neural network.

[20:12] which is hard, like we have to, [20:14] we have to kind of tell the whole body [20:16] how to reason over, [20:18] like camera frames. [20:19] and then how to like position itself fast [20:22] and dynamics. So you'll see in cases like logistics, [20:24] These packages are like moving. They're moving while we're grabbing them 'cause they're plastic. [20:28] We have to reason over all of that in real time. [20:30] and be able to do, like, you know, if things move in real time, we have to basically close-loop react to it. [20:36] That's like really hard. [20:37] We've been spending like the greater of like last two years trying to solve this problem. And I think [20:41] I think we probably have like some of the best AI policies for humanoid on the planet today. Sorcery is brought to you by Brex, the financial stack trusted by more than 30,000 companies, including one in three venture-backed startups in the U.S. Nearly 40% of startups bail because they run out of cash. Brex is literally built to help founders avoid that. Unlike traditional banks that let your money sit idle, chipping away at it with fees, Brex is designed to help you spend smarter and move faster. [21:11] Treasury and FDIC protection into one powerful account. You can send and receive money globally at lightning speeds, get 20 times the standard FDIC coverage through their partner banks, and even high yield from day one. With same day and even same hour liquidity, access your funds anytime. Companies like Scale AI, DoorDash, Service Titan, Hims, Anthropic, Flexport, Robinhood, and Plaid trust and use Brex. Start today at brex.com [21:41] That's B-R-E-X dot com slash sorcery.

[22:11] dot com slash S O U R C E R Y. Another thing I want to show you is, um, uh, this is like, um, [22:19] We just walked out of System Integration and Test, the lab, and we're trying to find all these different failure cases. One of our failure cases is for a humanoid, we're balancing. So if we lose power, [22:28] like the robot falls. [22:29] It's the same for like losing power or like losing a motor in the leg. [22:33] Do they need to be connected to Wi-Fi, too? What are all the... [22:37] - Yeah. - Integration. - These robots have 5G and Wi-Fi and Bluetooth, but we do not need to be connected to Wi-Fi to do work. Like these robots out here, if they lost like internet connection or network, they can basically continue to do work. We run Helix on board. [22:48] So they're actually loaded into GPUs onboard and memory, and we run inference onboard. Meaning if we lose internet connection, we can still do housework and logistics. [22:58] Like humans are. [22:59] - Yeah. - I mean, maybe I have a hard time doing work, would I lose another connection? But most humans can do most work. [23:05] So another thing that we like we're working on solving that I'm excited to show you here is [23:12] What if you lose communications with any of the 40 joints? [23:16] Or what if we lose power? [23:17] And upper body's kind of fine if you lose a wrist or elbow, [23:21] It's like, you're not gonna fall at the very least. [23:23] But falling is like a, it's, [23:25] terrible event for us. We don't ever want to fall. [23:27] We actually have an initiative internally called Never Fall. It's like we never ever want to fall. Even, you know, we will fall, but we don't ever want to tolerate it. The hardest problem here, [23:37] for the controllers, what if we lose like an ankle, a knee or hip?

[23:41] What if you just lost a knee? Normally for humanoid, you just literally fall over. You can't really balance if you lose a knee. We've been working on a project we call Vulcan here internally. [23:53] that basically allows us to lose [23:55] a single or even multiple joints in the legs and still not fall. [23:59] So what we're gonna do here is, Moritz, can you show her what would happen if we lost like a left knee? [24:04] So here we have a view of all the different joints on the robot. Green means we have comms and power to them. And the robot can basically communicate and it's fine. And red here will mean we'll lose certain communications with the robot. [24:17] So here we lost like a left knee, and you can see, the robot's kind of like hobbling on the left leg. So right now the knee is basically, we lost power communications now to the knee. The knee's locked. Got a locked knee. So we velocity lock the knee, and we can basically hobble around. [24:34] I thought it was gonna be a little bit more dramatic, honestly. It's not bad, right? It's not that bad. It's unbelievable that we can even do this kind of work. We're doing this also inside of a reinforcement learned [24:45] Neural Net Controller. So the same stuff that Morris talked about earlier, of us learning in simulation. We're about learning in simulation, how to move the body here. [24:52] to extend it lost different joints. [24:55] I watched the robot a few weeks ago. We were doing work on this basically logistics use case. [25:01] And it's like, you know, like, [25:03] Months ago, we were like losing knee, the robot would just fall. [25:06] Now, the robot loses a knee. It can either continue to do work or it can just hobble off. - Just limps. - Yeah, it can basically ask this buddy in the back, say, "I need another robot to come fill in." And the robot comes in and fills in and continues to do work.

[25:17] then we basically don't lose any time. [25:21] - It's pretty impressive. - Yeah, it's really impressive. I think this is kudos to the controls team. I think we have one of the best controls team in the world. About a year ago, we made a strong pivot away from code into neural networks. [25:33] And I think the team has probably shown like [25:35] I think some of the best neural networks for control [25:38] in the world on humanoids. - What happens if it loses its knee while it's bent? [25:43] It'll just, it'll basically do a velocity lock on that knee, and should be able to hobble around. It obviously depends where if we're like in a full squat down. Yeah. That might be really tough. [25:53] It depends what state it's in, but in most cases, we can basically recover from this at this point. [25:58] and survive. [25:59] You want to build a real-time operating system that is basically like fault-free. [26:03] So this team here is all focused on trying to figure out all of the potential faults and risks. [26:09] This team here is responsible for building our controller. [26:12] which is responsible for how do we move the entire, like all the joints on the robot. [26:16] keep balance and ultimately end up doing tasks. [26:19] How many different teams are in the company and how do you portion out who works on what and who gets integrated? This is part of our AI team. It's one of our biggest teams here. [26:28] We have multiple different groups inside the AI division. [26:32] We also have a team that has hardware, so they design motors, batteries, wire harness structure, joints, kinematics, like basically a wide range of stuff. [26:39] We have a platform software team that does all of our embedded software, firmware. We also have electronics team that basically builds like all the PCBs and electronics work that we do here internally. We designed almost a

[26:51] There's over a hundred PCBs that we design here on the robot that we do on that team. [26:55] things like our motor controllers, all this different type of work. [26:58] We have a system integration testing that you saw today that helps basically make sure we ship like really good robots out to the world. [27:04] We also have a team that does basically design, like industrial design, which we're going to show you towards the end of this tour. How do we design something that's like a... [27:16] Like people really want. [27:17] And I think we have a pretty high design standard here of designing something that is really delightful. [27:22] piece of technology. We also have a team that also work on system design and thermals. [27:26] The robot produces a lot of heat while it's moving around, and how do we get that heat out of the robot, and how do we design for it here. We also have fabrication teams that make fabrication prototypes. We have a Bocu team for manufacturing, supply chain team, [27:39] Um, [27:40] facilities, [27:42] And then we have like all of our business operations side. [27:45] So maybe like a dozen teams here internally that are needed to do this. [27:49] It's basically the same teams that you need to build robots. [27:52] So like any kind of robots, like when I, you know, at Archer, when I was building an aircraft, we had like, it's like a flying robot. So you have like the same type of stuff. You have like electric motors, batteries, control software. [28:03] embedded systems and sensors. [28:04] So we have basically teams around all that. [28:06] And then obviously on the AI side is a big focus here internally. [28:09] Do you want to see some stuff in the home? [28:11] Yeah, let's go. All right, let's go. Thanks, Moritz. [28:14] Thank you.

[28:18] I was hoping it would be more dramatic, but that's okay. Losing me? Yeah. Yeah, it's like, honestly, it's a really hard engineering problem that we're so proud of internally. We have a large initiative. Like one of the teams that we have here is like a Never Fall team. And it's a team that basically predicts [28:36] any potential faults in design around them. This is a case where like, I think you're always gonna have a period of time where you're gonna lose a lower body [28:44] like motor actuator. [28:46] then how do we survive this? Right. Okay. Outside of stuff we're doing on the... [28:50] commercial market, one of our big focuses here is how do we ship a general purpose robot [28:54] to do things like in the home. So come over here. [28:57] We have a robot here that is designed to tidy the house. [29:02] So clean up any [29:05] clean up the table, like basically put away all the different cups, clean the toys up, [29:11] Tidy the couch. [29:13] Like, you know, things in my house is kind of chaos. - Oh, it's spraying and cleaning the table. - Oh yeah. [29:20] Don't you want, we all need this. It's like, [29:24] That was a little sassy. It was. [29:28] So what's cool here, [29:30] is the robot is running-- [29:32] an onboard Helix 2 neural network, [29:35] to be able to do all this work, to be able to do all this cleaning. [29:39] So it's basically just taking a prompt, which is like clean the living room. [29:44] And it's basically reasoning through what to do from its cameras. [29:47] and it's ultimately telling the whole body what to do from a neural network. From an AI policy. How many hours

[29:52] has this been trained on? Doesn't it take millions of hours to train? We've probably had in Helix, like the like in like kind of base pre and mid training, [30:03] Helix models that we're working on now have [30:06] Maybe like a little under a million hours of total data. And then we also do from mid training and pre training, we do post training, [30:16] Here, which probably is I would say low like basically thousands of hours that are running here The goal for us is design a single [30:25] kind of neural network platform. [30:28] that can basically do things like tidiest living room, but also do things like logistics. [30:31] And this isn't tele-operated? This is not tele-operated. There are rumors that these are tele-operated. For sure not tele-operated. You don't have a secret room back here? Who's in there? No secret room. These robots are running purely autonomously from onboard [30:46] AI policy called Helix running on board the robot in the torso and so this robots job is to clean this room up and does this robot do this over and over again each day or do they take turns like how do you any robot in the fleet can do any of this work here. Okay, so it all connects. [31:02] Yeah, it all connects. We run a single neural network that can basically run. It's the goal. It's the reason why humanoid is so great. It's like the same humanoid can go over and do logistics and healthcare and manufacturing or do dishes, just like we can. Right. So our platform here is to basically run a single neural network that we call Helix on board the robot that can multitask between different things.

[31:23] When they're in people's homes, will that [31:27] data still be trained and will it stay local or will it be spread throughout the network? Yeah, we basically, we need data, basically the biggest blocker for us now of going from where we're at today to like large-scale deployment is data. [31:42] We need like an enormous amount of data. [31:44] We need to pull a lot of our resources further into pre-training for the Helix team. [31:49] And we just need a lot of diverse, really high-quality data across the world. This means data in the home, means maybe data more in the commercial market. So we have two efforts going on here. One is basically a large-scale data collection effort that we're doing now at the company. And two is when we deploy robots, we do want to be collecting data, and we do want to be training on that. [32:05] And we do want to be sending that out up, training on it into a central training jobs. And then we want to be software updating the robots with the latest neural network weights. Does it get anonymized? How do you deal with the privacy? Oh, yeah. We like fully want to anonymize all of it. [32:19] There's a lot of data we really don't care about. Most of the data we care about is like, what's the robot from a state perspective, like scene? And how do we use that to basically train the robot to be more, like, to generalize better in the future at those different areas? Are you sticking mostly to the U.S. now? Because if you go to Europe, there's obviously a lot of... Most of all of our work today is in the U.S. Okay. Yeah, we do want to be global, though. Yeah, Europe's a little tricky. [32:42] Yeah, what are your plans? How do you skirt around their data privacy? Basically, we got to play by the rules in Europe. [32:50] Our hypothesis here is that we're missing a certain set of data that we're collecting now.

[32:56] that will allow the robot to generalize in almost every condition it sees. [32:59] We're kind of going off and doing the same things every day. We're doing dishes and laundry and tidying the home. Like the same stuff we're seeing. We're kind of doing the same movements, like grabbing some off the ground, putting it away, or pulling the dishwasher open. Like it's kind of the same stuff. [33:12] Our hypothesis now is that we will see enormous amount of positive transfer from the data collection efforts we're doing now into pre-training across like [33:19] basically any environment in the world. [33:21] I mean, that'll be like, it'll be some way, you know, we'll approach that somewhat like, it'll take, I would say, [33:29] over time, a large amount of data to find every out of distribution. Basically, be able to do everything possible in the world, but we think there's a path to do this. [33:36] - And what is the price point differentiation for the in-home robot versus commercial? - The in-home, we're not selling right now to the home. We want to sell here in the near term. [33:49] and we want to sell the robot for hundreds of dollars a month. [33:52] And as a, like, somewhere like a car lease, maybe like four or five, 600 bucks a month. [33:57] How are you thinking of deploying them in homes? Like, do you have to see if people have enough room? [34:02] Like in New York City, I can't imagine these little apartments. It takes a dock, it's like two feet by two feet. You can plug it in a wall outlet, it'll go to a stock and charge. And then throughout the day, it'll just go off and do work. Whatever you want it to do. [34:16] Like for me, I wanted to do like the laundry probably almost every day. [34:19] dishes every day and tidy the house. [34:21] multiple times a day. [34:23] Do they have a distinctive diet? What do they eat? They eat nothing. They're keto? They just work 24-7. They're just constantly intermittent fasting. They're in this, like, they're in this, like,

[34:32] They're in this like purgatory state of just working 24/7 for us their lives. [34:36] Wow. [34:37] Yeah. [34:37] FUN. [34:39] Okay, I want to next go show you how we make them over at Baku. Yeah, great. Let's go. [34:44] It's you know like kind of like HQ, but Bacu. [34:48] - It's the HQ for bots. - Oh, it's BotQ. - Yeah, BotQ. - I thought you were saying Ba-ku. [34:53] - No. - Isn't BotQ? - BotQ. - Botchy, no, anyways. - This is RobotQ, this is Robot Quarters. [35:01] One thing that's really cool is on the way is we work at BMW. And last year we deployed for six months robots on the basically body shop factory line to build cars. Yeah. And this was the first. You built this entire car? No, not the entire thing, but we helped build this car. Okay. This is an X3 that we helped basically, like we basically helped, the robot helped assemble it. And this was the first car in the world built by a humanoid robot that we're aware of. [35:31] - Straight off assembly line? - Straight off assembly line. I actually, I bought the first four. We have three here on campus and I have one at my house. And yeah, it's like, [35:39] Collector's item now. That's exciting. Yeah, it's pretty pretty interesting. I [35:42] All right. [35:45] So we have, this is kind of our campus here. We have four buildings. We're gonna go through our manufacturing site, which is BotQ. And then we also have a site up here that I'll show you, called The Grid. [36:00] And the goal of that facility is to run robots, like just like we would at our client sites. Could be in the home, and also on the commercial side, and 24/7 operations. Why is it called The Grid? It's a kind of a nod to like a sci-fi movie. And...

[36:17] You have a lot of inspiration from sci-fi movies. Are you a sci-fi geek? It's a total sci-fi geek. I've seen every sci-fi movie. What's your favorite? Probably "Contact." [36:28] Jodie Foster. - The alien thing? - Yeah. It's kind of embarrassing to say, but-- - Why? - I don't know. The contact's amazing. - If you like it, don't be embarrassed. - Yeah. But I'll just, I'll watch, I'm a sci-fi junkie. I'll watch any sci-fi. [36:43] Yeah, so that's the grid up here. And we'll run robots in that facility 24/7. And the goal of that is last line of defense before we send out any code to our customers. [36:56] So you don't want robots to... [36:59] You know, you don't want robots having any problems. We want to run them close enough to heavy operations that we would see out in the real world. And so we have a whole facility dedicated to basically running robots as hard as possible 24/7. We run on holidays, weekends. [37:13] 2 to 3 in the morning, they just run all day, every day. - Have any escaped? [37:17] We've had one almost escaped. - Really? - No, nothing's escaped. - Do you geofence them within properties? Do you do that when you set them up? - We track them. Obviously this is like for us right now, these are like high IP, very complicated hardware. We don't want to get stolen or out in the wrong hands. [37:35] So we do track it. Has that happened before? Are people stealing your IP? We do a lot of work on security internally here. So we haven't... [37:44] had any known IP thefts. [37:46] That's company. - Okay.

[37:48] All right, welcome to Baku. [37:51] All right now, oh my god now welcome to buck you So this is where we this is where we manufacture our [37:59] Figure three robots. [38:01] - Wow. - So we do everything from build heads, batteries, [38:06] legs, arms, fingers, thumbs, hands, [38:10] And we basically do all testing here before we box the robots out or they walk next door. What does a box look like? Box? We'll show you we have one over there in a minute. Really? And right now we basically if we need them at headquarters of the grid They basically just walk over. Where do you store them? At the office or client sites. [38:27] Yeah, on the docks. They basically dock at nighttime. [38:31] or whenever they're not needed. Okay, we're gonna show you some of the manufacturing lines. So we start first with basically head and battery. [38:40] and we do some electronics, basically quality and EOL checking. [38:45] EOL is end of line, so we make sure we want every single subsystem to go through a pretty crazy test. Here's our headline. A rack of heads? Rack of heads. So... [38:55] Here. [38:56] Here's a head. [38:58] Our heads have [39:00] - Yeah. [39:01] basically Bluetooth [39:04] like Wi-Fi, 5G, they have camera systems on board. We have lights, we have thermal systems. [39:10] We have an IMU. So basically the head is like basically a lot of sensors. [39:14] in here. [39:15] The heads go through a pretty [39:18] like a rigorous test.

[39:20] Here [39:22] that we've designed internally for end of line testing. [39:26] So heads here go through, well first is we basically our flashing software here. [39:31] under the head for the first time, [39:32] all the firmware, [39:34] It's going through a calibration process for the cameras. [39:37] and then we're basically making sure their head is in a nominal state to basically put onto a robot. So are we getting signals out of it? [39:44] Does it have any issues at all or any errors? If it does, we'll try to triage it. If it doesn't, we'll end up putting it on a robot. [39:51] This is like when it's first born. [39:53] Yeah, it's like, it's kind of just like just raw hardware and it goes in here and it comes out with software and comes out with all the checks that we can use it with. [40:00] Wow. Yeah. [40:03] All right. [40:06] How often do you walk through every part of the campus? Every day. [40:13] So here is our battery line. [40:15] So we have battery cells that come in, [40:18] and then we basically do cell testing and basically voltage balancing. [40:23] So here we're basically checking every single cell against the data sheet. [40:27] We're also checking voltage and we're balancing out the packs. [40:31] So is any-- [40:32] a kind of voltage differential. We're basically making sure that all the packs basically have somewhat of a somewhat balanced voltage. - Where do you get this machine from? - We designed this machine. - Oh. - Yeah, this was custom designed. [40:44] for figure here, for battery. [40:47] And then we go through a process for potting, wire bonding, and there's some polyurethane we put inside the battery pack for thermal runaway. And then at the end,

[40:58] Pops out basically a pretty heavy 2.25 kilowatt hour battery pack. [41:04] Yeah, I don't even try it's really quite heavy. Oh, yeah [41:11] No, that is actually So this uh, so this battery will basically go right into the torso and [41:18] It's one of the heaviest components we have. [41:20] - Yeah. - How do you-- I mean, I know all this is stabilized, but, like, [41:24] Is it better that it's one piece in the torso versus distributed across the body? Yeah, it's way better one piece. [41:30] The battery pack has, just even for safety, we have a lot of thermal runaway properties inside the pack that we've designed here internally to make sure-- in the worst case, you basically want [41:42] to say like, okay, if a cell or battery cell is going to thermorunaway, you never want that to ever propagate outside the pack. [41:48] So you wanted to contain it to the battery system itself. [41:51] So we have basically a structural system and also a thermal and away venting process we've designed internally to allow for the battery to be extremely safe. [41:59] Like the requirement is like you want no flame to ever exit the pack. [42:02] You don't want a robot like [42:04] on fire or something like that out in the world. So we've designed the right-- - Have any of them? [42:08] We've never had a robot ever have this. [42:11] - No. - No. - And then all of our figure threes are designed in a way that basically will prevent [42:15] the robot from Ever Catching Fire. So that actually was a pretty crazy hard engineering feat that we designed here internally. [42:23] Also, the pack is structural, take loads. So in case we fall, [42:27] even like sharp objects or corners and things like this, like we can never propagate inside the pack.

[42:32] the cells itself, meaning you don't want anything to kind of like, kind of like send the battery cells itself into thermal runaway conditions. [42:40] Have you had any supply chain risks? [42:42] whether it's with China or other countries, is that why you do everything here? - We do most of the manufacturing here because [42:48] The product is so new and it needs to be really controlled. And we also need to, we also think about IP as really important here. We don't want any of the technologies to be stolen. And this is just hard too. Like we'd be able to put this thing together through a brand new supply chain that we had to design. And get it in and make it work is like, it's non-trivial. Like you see how much testing we do at headquarters for all this work. [43:08] testing we do with the grid. [43:10] We'll do a ton of testing we'll show you here today. It's enormous and the product is [43:15] We're kind of like early in the humanoid [43:17] kind of like chapter book. So like cars have been around for [43:20] Over a century, we kind of-- - The company's been around for four years, no? Yeah, not even four years yet. [43:25] Yeah. And then humanoids are like really early in that whole process. [43:30] How did you? [43:31] We'll talk about this more in the long form, but like... [43:35] In terms of getting this up to scale so fast, like you have now created a humanoid robot. This is one of the most complex. [43:41] Mm-hmm. [43:42] robot [43:43] like, [43:44] I don't know, engineering problems ever. So like, how did you get up to speed so fast? [43:49] My company before this designed flying robots at Archer. [43:53] And it's got the same properties. We have a battery pack, but instead of like two kilowatt hours, it's 160 kilowatt hours and it's distributed. [44:00] We have electric motors, we have control software, we have embedded systems and sensors.

[44:04] That's a robot. And Archer's aircraft, my aircrafts there at, say, midnight, are highly overactuated. [44:11] All the propellers have variable pitch. [44:14] The front leading edge actuators tilt 90 degrees. You have the flaps on the wings and tail all move. You have like basically 24 degrees of freedom. About a little over 40 here. [44:24] on the robot. So in some way it's like, [44:28] similar enough systems here. And then, you know, when I started Figure, we were like, we have a very crisp and clear vision for the [44:37] "How to think about the product and engineering roadmap?" And we just went like, [44:40] went really hard. [44:41] building a team to [redacted address] for us to design stuff really fast. We have the, we'll show you here next visit, but we have our figure one robot. [44:50] It's gnarly, it's got wires everywhere, it's our first generation. [44:53] We had that walking before we were a year old. [44:56] We think it's probably one of the fastest times in human history. So it was just like a... [44:59] You know, we were like laser focused, pedal to the metal, trying to get this thing to work. Yeah. Yeah. Okay, let me show you some more stuff. So we have a bunch of different lines here. [45:08] that helps build pelvises, [45:10] install battery, compute arms, legs, [45:14] Here we're installing the lower leg. - Is there a reason there are humans installing the leg? [45:19] At some point we will have robots doing all of this work. Don't let them hear that. Yeah. And we're putting more and more automation in lines now. We will be shipping our humanoid robots into the production lines here. [45:33] this year and

[45:35] And yeah, now we're doing the lower leg assembly for this robot. [45:38] And at some point today, [45:39] It will go through some testing, we'll show you in a minute, and we'll basically walk over to headquarters and start basically helping us either do AI development or doing use case testing for our customers. [45:52] Pretty cool, huh? - It's pretty cool, yeah. [45:55] - Yeah. - It's so crazy seeing them get assembled. [45:59] Are you worried they're going to become sentient? [46:02] Um... [46:04] I... [46:06] I think we'll be okay. [46:09] Um, [46:10] These things will get really smart. They're able to do what humans can physically. I think the neural network technologies we're designing are trying to give human common sense to all the robots. [46:21] So in some way, I think we'll get to at or even beyond human level intelligence in these systems. [46:27] It actually might be the case that we [46:29] We get to artificial general intelligence first in these embodiments. [46:32] Really? [46:33] - Why? - Because we're able to, this interaction data, like touching the world, and seeing what happens through trial and error is like, it basically, most of human intelligence, [46:44] is built this way. [46:46] And I think this is the last missing piece to get the true AGI, is this real world interaction with our environments. [46:54] Okay, come with me over here. I'm gonna show you some of the testing we were. So the robots are basically built [46:59] starting with the pelvis, torso, head, arms, legs and hands. [47:04] And then we basically want to go through a very strenuous testing process.

[47:09] to make sure everything is working nominally before we send it out. [47:11] So we don't want any loose cables or bad parts or bad communication. So we send the robot through what we call a final EOL test or end of line test. [47:21] This is where the robots go through all the final checks, and they basically go through also a burn-in. [47:25] on these lines. - What do you mean by burnin'? - They'll run for several hours. And we'll basically make sure there's nothing that, no issues pop up over those few hours before we send the robots out. So here we have these bays that are running [47:42] a combination of burn intestine, [47:44] and end of line checks that we've designed here internally. [47:47] So we basically go through a process where the robots basically self [47:50] like trying to understand itself, like, does anything seem wrong? If it is, we will flag it. If it's not, basically, we go through a process where we basically do, like, basically a bunch of checks and burn in, make sure the robot's in good condition. [48:03] If they pass here, we basically will walk over to headquarters. [48:07] And if they fail here, we need to go fix it and understand why. [48:10] Why do these ones have vests on them? [48:12] - When the robots are getting brought up, we basically, I think we hold it through a gantry system on the back. [48:19] and they have their vests on for that system. [48:24] It's basically like the robots are like, [48:25] just been born and they're waking up, they're saying, [48:28] They look at their hands, they start calibrating itself visually, [48:32] And, um, [48:33] trying to make sure everything's in a healthy state. [48:36] Of this campus, what is your favorite part?

[48:40] Thank you. [48:40] Thank you. [48:41] I think Baku is one of my favorite parts here. Like being able to build robots [48:46] In March, we had record [48:48] We had record manufacturing. We made more robots in March than we had ever in our entire lifetime combined. It's just cool to see us be able to do this and then get them out the door. [49:01] So I would say this is probably one of my favorite places in the campus. I think maybe my other favorite place that will show some point here is like, the robots doing really useful work. [49:11] 24/7. [49:13] on either commercial customers or in the home. That stuff is just amazing. [49:17] Because what we're here to do is we're here to basically build a human-like [49:21] intelligence in the world. [49:22] And to see robots working 24/7, being able to do things like human scan is so special. It's such a hard thing to do. And be able to see us doing it, at these levels of reliability, it's awesome. So I think it's probably a couple of my favorite places on campus. - If you didn't have your job, [49:37] What job would you want? [49:38] Here? Yeah. I think [49:44] I think there's a few things I really like. I like the engineering design process of how do you think about clean sheet, improving the system to be like more reliable, cheaper, [49:52] lower in mass and overall a better functioning robot that can do more of what humans can with less complexity. [49:58] I think that job across the hardware engineering [50:02] in Software Engineering Design Org is like, I spent a lot of time in that. [50:06] like leading engineer in here and it's just like a [50:09] Yeah, it's just a really fun

[50:11] and I think very hard problem. [50:12] Every choice you make to try to make the robot better for thermals or lower mass or lower weight, makes something else on the other side worse. [50:22] Like if you're trying to make the robot lighter, it means it probably can't hold as much weight then. If it can't hold as much weight, the customers are like, well, if you can't carry 30-pound boxes around, I can't use you on this assembly line. [50:33] So there's just like a, I think there's a lot of interesting, very hard problems to solve there. I think second is like how do we get neural networks? [50:40] to run on robots and generalize at scale on the Helix team. And that really is, at this point, a data and generalization issue. And that's a really hard, fun problem. [50:51] Manufacturing is another one of how do we manufacture robots at scale? How do we continue to get robots in the manufacturing process to build themselves? [50:58] and how do we get them to off the lines into the real world as fast as possible? At some point here, it'll just be like a... [51:04] It'll just be full lights out manufacturing. [51:06] We'll have robots only building robots and sending out to the world. Robots will be getting into boxes themselves. Other robots will be boxing them up. And we'll be shipping them out to customers. [51:15] - Yeah. - Sounds a bit sentient. - It's a bit sentient, yeah. And so that's another area where I'm just like, it's extremely interesting. I think the last piece is, [51:24] Um, [51:25] We have a whole commercial operations team that's trying to get robots out to the world at scale. [51:29] and make them really useful. We did this with BMW last year, and we're doing it with more customers this year for Figure 3. [51:35] And it's a really cool problem because it's really hard. [51:38] Like robots need to get in the environment, need to get safely. They never, like basically can almost never fault.

[51:42] And when they do fault, they need to understand that and self-correct. And then we need to be able to do useful human work at human performance. [51:48] So our comparison is like, what does a human do today? In terms of speeds and accuracies. [51:53] and then reliability. So like that's a hard bar to hit. [51:56] So those are all like kind of like gigantic problems. [51:59] to go solve that I think if I was here, I'd want to spend all my time on those. [52:04] - Oh. - Yeah. [52:05] Awesome. [52:06] Okay, you want to see the design studio? [52:08] Yeah, let's go. Let's go. [52:13] What do you think of this? [52:14] It's really fun. [52:15] It's really fun. [52:16] I have the coolest job because I just get to visit people's factories all day. Yeah. And everybody's building something different. I was at Applied Intuition earlier. [52:24] This week I was then at Skydio, I've been to Archer, I've been to Anduril, [52:33] We have just [52:35] Been looking through every door. That's awesome. Yeah. [52:41] VCX by Fundrise, the public ticker for private tech, allowing investors of all sizes to invest in venture capital. View the portfolio at GetVCX.com. That's GetVCX.com. [52:55] Some of you may not have heard this yet, but our sponsor Public just launched something called Generated Assets, and it brings AI into investing in a way I've honestly never seen before. Here's how it works. You type in an idea like AI-powered supply chain companies with positive free cash flow or defense tech companies growing revenue over 25% year over year.

[53:25] why each stock is included. And before you invest, you can even backtest your idea against the S&P 500, [53:31] So you're making decisions with real context, not just guessing. [53:34] And beyond generated assets, Public lets you invest in stocks, [53:38] bonds, options, crypto, all in one place. They'll even give you an uncapped 1% match when you transfer your investments over from another platform. If you want to build a portfolio that actually reflects your thesis, visit public.com/sourcery. [53:51] Paid for by public investing. Full disclosures in the description. [53:55] Enterprise AI runs on Merge, the AI infra platform for integrations, agent tooling, and model orchestration, so your teams ship product, not plumbing. [54:04] Mistral, Dropbox, and Drada already trust Merge and production. Start building at merge.dev. [54:11] Founders scale faster on Deal. Set up payroll for any country in minutes. Hire anyone anywhere. Get visas handled fast. And get back to building. Visit deal.com slash sorcery. That's D-E-E-L dot com slash sorcery. [54:27] - Five years ago when I took Archer Public, I saw this campus, it was all empty, and I was like, "I gotta be here." - Okay. - The buildings are really big. It's just like they're all open. It's just like multiple buildings on campus, and they're all close. And it's hard to find space like that where you can build hardware in, make loud noises, but also close where everybody lives in the Bay Area. - Are you worried it's too nice? [54:48] - No, not at all. - No. - It looks like we're manufacturing in here and running robots, so it's fine for us. It has some industrial grit here.

[55:01] You know what I mean? It feels like we're kind of like builders and makers. So it's kind of nice. It kind of reflects a lot about who we are [55:09] on the team and it's just like my happy place. Been looking at this for five years, we finally got it and it's great. And I can just walk to manufacturing, I can walk to the grid, I can walk to building four, I can walk here to headquarters and I can just be here with my team and pop in where I need to solve like any, like whatever the biggest problem of the day is. [55:30] Is most of your talent down in South Bay or do they all commute? Most of the talent is here and then we have a shuttle for folks in the city that want to come down. So maybe like, I don't know, 5 or 10% of the folks here live in the city and commute down. And then I think maybe the majority of the rest all live kind of pretty local within like 20 or 30 minutes from the office. [55:49] - Yeah, luckily the weather is nice here today. - The weather is just-- - Freezing and cold and rainy yesterday. - Yeah, yeah, the weather's pretty much the same here every day. [55:58] Okay, let's-- - Oh my God, we're back. - We're back. So fast. [56:04] How did you come up with the logo? Okay, the logo is, I've got a couple things that are cool here. It's like one is like, it kind of like looks like, like basically how the robot steps. And tracking its steps and feet. Yeah. And then two, it's like a little F. [56:19] For figure... [56:22] So this is the fancy secret room now? [56:24] This is a secret room that nobody's allowed to come in. [56:27] This is our design studio. So I'm going to show you every robot we've ever built.

[56:32] Wow. Yeah. [56:35] So we started in 2022, and the goal was, like, how do we get humanoid robots to our AI and software team as fast as possible? [56:43] So we designed figure one here. [56:45] This is our first-generation robot. [56:47] Uh... [56:49] Pretty cyberpunk. [56:51] How much did this one cost to make and develop? [56:56] going down the line. [56:57] Oh, wow. This one was like built to be expensive and move extremely fast in terms of building it. So this was like hundreds of thousands of dollars. And the robots we have now are like, you know, well under $100,000 each. [57:17] And, uh, [57:18] So yeah, this was very expensive, mostly expensive because we CNC manufactured the entire thing. [57:24] Basically the way we made all the metal parts was like extremely high precision like things like Formula One race car type type stuff We had this walk-in within the first year And we did a lot of the early AI stuff here that we kind of proved out the company was it was great And then we moved on to figure two which is which is here some improvements we did as we moved the battery and [57:46] that was on a backpack. [57:48] into the torso. And then this one had like a basically relatively small computer compared to this. We basically tripled the compute. So we basically doubled the battery pack. [57:56] Triple the compute we have new camera sensors in the head and pelvis in the back we had our next-generation hands and

[58:04] And we basically wired the whole robot internally. [58:07] And we also designed the structure similar to aircraft. Aircrafts take loads to the aircraft skin. And so we designed the, basically the structure is an exoskeleton. So all the load paths [58:20] uh for the structures is is exterior of the exterior surface um [58:25] We made about, I think like 50 of these. [58:28] And we just recently, we're just, we're just, we just recently tired about a month or two. [58:32] And then we've moved now to generation three. [58:35] This is our figure three robot. Those are all three the same with different outfits and The robot here we like reduced we reduced the weight we made it skinnier, but also keep the same power and torques It's got a Zimpic. Yeah, exactly [58:51] It looks better like this look a little too like Kind of too... Roboty. Yeah robot less robot too much robot. Oh [58:58] Exactly. So we slimmed everything down. We soft wrapped it. It's got a layer of foam on the shoulders and the chest. So it's soft. We have our newest generation hands on this, which have [59:09] camera, tactile sensors, [59:12] It's basically able to grasp items much, much easier. We reduced the cost by about 90% between these two. - Really? - Yeah. - What was the major cost there? - We didn't care. We've optimized these first few generations for speed. [59:27] We didn't care as much about cost. So that was like the biggest misconception of designing it initially was speed? Yeah, like we were sitting here in 2022 and we're like,

[59:37] Um [59:38] our software folks need a humanoid to do testing, and do AI work, or whatever it is on it. And there wasn't a time, still really isn't, a good humanoid robot to go by to help us speed us up. So we had to go build it. [59:51] Even if it's expensive, let's get stuff to the team as fast as possible to start getting like the, basically start working on the development process for commercialization. And the same with figure two. The goal was like we had a lot of problems with reliability here that we needed to clean up. We had like wires poking out and all kinds of issues here with this robot. It was kind of faulting every few hours. So we just had a ton of reliability issues we needed to clean up. Because we're kind of known though because we're moving really fast. So I had figure two. It was just like way too expensive. Like too hard to manufacture at scale. [1:00:21] So figure three was like, how do we reduce the cost by almost an order of magnitude? How do we make a lot of them? How do we make it closer to what we think the ideal outcome is for every robot in the world? So these robots basically have the ability to take these clothes on and off. So we have different types of accessories we can put on them. Shoes, gloves, like... [1:00:43] Yeah. [1:00:44] - How often are you thinking about the next version? Like what do you want? [1:00:48] the next version to be and like, do you start thinking about it while you're building this one or after it? Like at what point of the production stage do you start the next iteration? We are now building a new robot almost every year. [1:01:00] And we'll have, you know, we were like, we're like late stage now in the design process for figure four.

[1:01:07] And... [1:01:09] There are some changes that we want to put into the iteration. This is like iPhone sale, like iPhones. You know what I mean? Everyone's getting better. This is Apple. Yeah, exactly. So everyone's trying to get better. There are some things we just didn't get to and have enough time to work on, on figure three, that we want to de-risk. [1:01:25] We put in a four. [1:01:26] And there are also some things we've learned from like operating figure three now that we're like, man, this is like could be way better if we did this or that. That we're putting into figure four as well. And then we want to keep reducing costs and making it easier to manufacture at every step. So we're looking at figure four. We're like, figure three, we're like, oh, man, like some of these things are really kind of hard to manufacture at Baku. [1:01:45] Or how are we going to get it out of the box? Or how are we going to get a new user in a home really easily? So we're taking all those collective learnings and we're putting them into the engineering design efforts. And we go through many different basically gating processes for that, starting with an architecture review of what the system should look like, and high-level level zero requirements, all the way through to detail design, which we're now in for figure four. What's kind of crazy here is I thought at some point we would saturate out. You know, iPhone doesn't really change much anymore. [1:02:15] I thought like figure three at this point was like, this is like our best human robot in the world. And every robot's gonna be like, it's gonna be better, but not by much. [1:02:22] What's going to happen here is you're going to have figure one, to figure two, a step up. Then figure two to figure three had a step up. Figure four will be the biggest step up we've ever made. [1:02:32] by far. We'll have it out here at one point and you'll be like, oh my gosh, it's just like radically different. And, um, so we obviously can't talk too much. We can't,

[1:02:40] talked about anything basically on what we're doing there but like uh we are just so early we're like almost in like flip phones and now we're entering like iphone one moment i think maybe figure four will be our first like iphone one moment for this where it's just like radically different and probably like [1:02:56] For me, I think it's like almost the perfect humanoid robot I can think of. I'm sure there's things on five and six as we iterate through it'll be even better. [1:03:04] But we've learned a lot. Here's a couple examples. We have, here's some parts we have on the table for, this is our generation of hands, starting with our first generation hand to our current generation hand today. We've gone through like five, [1:03:18] versions of this. - Well, they're kind of like the same size as my hand. - Yeah. One thing we have never shown is our first-generation hand. - Really? - Yeah. - Why? - It was really difficult from an IP perspective, an engineering perspective, to go build. And two is we think we learned a lot about it and we learned like things like what are good and bad. In this case, we felt like we had learned a lot about the hand of like why it's probably not the right direction. So our first-generation hand you can see here [1:03:45] is a tendon-driven hand. We designed all the motors and actuators here ourselves. [1:03:50] even the gearboxes. [1:03:52] Basically the rationale here is that a human hand is like this. Most of our motors are in our forearm and they're basically, we're like little tendons basically driving all the fingers. And so our first generation hand is like how do we get a really dexterous hand built, which is like really good for intelligence and AI. [1:04:09] And then ultimately, I can do a lot of things that human can. And how do we mimic the biological architecture of a human?

[1:04:16] And [1:04:18] So I was like, this is going to be great. We're going to put motors in here. They're going to be really powerful. They're going to drive a really high degree of freedom hand. And it ended up becoming the wrong engineering choice. And we ended up pivoting away from it really early. Most of the wrist motors are also in here. So on figure one, I don't know if you noticed, but the wrists look crazy. And the reason for this is that we pivoted away really early away from this tendon driven hand. We had to figure out a way to get new motors in for the wrist. [1:04:45] So instead of waiting like four months to redesign those from scratch, I took the motors from the feet. [1:04:52] So we have three foot motors here in the forearm, and it's just like this Frankenstein forearm. And it bends like, instead of bending here at the wrist, it bends like halfway through the forearm, which is just like really weird. And I was so ashamed, I'm like, we're gonna get this thing out, [1:05:08] At the time, I was like, this is incredible. Did you sell any of these? We didn't sell them. We just used them internally. [1:05:13] But we showed it and it was like this big like forum. I was like everybody's gonna notice this big forum is so weird Yeah, and I don't think I've ever had a single person in like three years ask why the forum was like this [1:05:26] It's not a robot not a single one. Yeah, so we uh, [1:05:30] So we ended up pivoting away from like the tendon driven hand [1:05:33] to our current generations of hands, [1:05:35] And, um, [1:05:37] We've learned a lot, but like I, yeah, we ultimately I think are building like some of the best hand technology in the world. We recently unveiled our high degree of freedom hand as a teaser, our next generation hand, about a month ago.

[1:05:52] This hand has basically a human level dexterity. [1:05:57] like as many joints in the hand as a human. And this is really important not just for like being able to do this task, but we need to be able to learn passively from humans at scale. And if humans can move hands in all these different crazy way, we need to be able to map to this. [1:06:10] at test time on the robot. [1:06:12] So we have, I think this is extremely important to get, if we want to solve like [1:06:16] uh AGI and get to like human intelligence in the physical world like it's all going to start here with the hands for us. [1:06:22] Yeah. [1:06:23] - It's intense, it's complex. I did see a couple people walking around the campus in spandex outfits. - Yeah, it's a mandatory outfit. It worked for me. [1:06:35] Yeah, like you just gotta be in spandex so you can't come in. [1:06:39] Yeah, so we basically are doing a lot of data collection here where we're trying to basically do like joint level tracking and different type of data collection efforts, like learning from humans. Like our training set is like, how do we like we're humanoid. We need to learn from humans at scale. And so we're trying to learn as much as possible about human movements and like image conditioning these policies here at Figure. Is that the oddest job? [1:07:01] - What is that called? - What is the oddest job in the office? - Yeah. - You think that's odd? [1:07:08] I think it's kind of cool. - It's uncommon. - Yeah, it's uncommon. [1:07:13] Probably the honest job we have there. [1:07:15] Yeah. How does one apply? You apply on the site. You apply on the website. We do like, we basically have like data collection folks that we basically are here that we have both here and out in the world doing data collection for Helix.

[1:07:27] That's cool. Yeah. [1:07:29] Yeah, I actually think it's a really cool job. Have you ever tried it? Full spandex? Yeah. I haven't tried the full spandex, but I've done every other type of data collection effort here. Maybe that'll be your next job. Yeah, I'm gonna go get in some spandex later and I'll mess it out. Okay, so we saw the generation of hands. Okay, so generation of hands. We also have some mock-ups for the head and feet. I think the feet are actually really interesting here. [1:07:52] This is kind of our first prototype for figure three. It's like basically like we really wanted to get a tow. [1:07:58] in the robot, which is important both for like a natural looking gait like as you walk, but also like getting off off the ground. Yeah. It's really important or even squatting down. When we squat down we're like on our toe box. [1:08:08] and it's really difficult with a flat foot. And then basically this is our, this is our figure two, like foot, it's basically just a, [1:08:15] fixed piece of metal. Nothing too crazy impressive. And then this is our current generation of fly nets. It's for figure three. We basically have a tow, [1:08:26] We have this opening here in the foot. Some people ask about this. This is basically for thermal venting as we're charging. We're pushing air through the calf and the shin, through the foot to cool it down as it's charging 'cause it's got inductive coils on the bottom. So these feet are basically stepping onto the charger, [1:08:47] We then initiate charge wirelessly. [1:08:50] and the robot can charge at two kilowatts. [1:08:52] So basically we can charge for an hour. [1:08:54] by standing there. [1:08:55] That's really cool. That's really cool. Yeah, we can charge like at client sites like this. We can walk over. We can dock to it. We can just stand on it. Over time, we're going to get these systems even smaller and be able to put them, you basically be able to put it anywhere and you can plug this into a normal wall outlet.

[1:09:14] Charging are you gonna do any brand deals? Um like foot brand like foot deals. I would love to do maybe not foot But your deals if like if Nike's watching like you're gonna be our next Yeah, Nike sneaker dealer. Yeah, I really wanted a high top for figure three. I think it's just like so cool Yeah, it's like our figure two looks like a like a penny loafer and just like you know what I mean? Can't be having that no can't have a penny loafer out here. It's like this high tops like kind of made of doing work. So I [1:09:38] Of the designs here, you have quite a sleek design. It's very futuristic. [1:09:44] How did... how did... like... [1:09:46] How did you get to this point? Yeah. We have an industrial design team here internally run by David. He just met. And we have a team that are obsessed with trying to create like a – yeah, trying to create like the – [1:09:59] Um [1:10:01] We want to create something that... [1:10:04] is really delightful to be around. [1:10:06] And it's not just like the way the robot looks or the size of it, it's how it walks and interacts and how its body language and how the human machine interaction is. [1:10:16] Does it look at you while it's talking? How do we deal with speech? What do we do with like, do we have like three screens in the head, like what do we show there? How do it make us like really, [1:10:24] Pleasant to be around are there any I mean you love sci-fi were there any sci-fi movies where you wanted? Oh, yeah, like the robot movies I robot ex machina like which ones? I think a thing we always talk about here is like there's like two roads for humanoids There's like a road to head like down like the robotic road, which is like I robot and [1:10:41] and there's a road to head down for like Westworld. [1:10:43] - Okay. - It's also humanoid.

[1:10:46] Where do we go? [1:10:48] What do you think? [1:10:49] Um... [1:10:50] I mean, I'm a big fan of Ex Machina. Okay, so what we go to Westworld. Yeah. Yeah. Okay, let's do it. [1:10:57] Yeah. [1:10:58] We're heading to Wesselworld. I think lastly is we're super proud to be on the cover of Time Magazine this past year, which was really cool. We had a robot in a home basically doing like [1:11:11] full like, you know, like doing housework with Helix AI system that we designed here internally. What's with the Deadmau5 record? Yeah. [1:11:21] We had Deadmau5 at our holiday party two years ago. And he was like, this is insane. We had robots on stage. He's like, I got to get you guys out to concerts with me. And we got to be like dancing on stage. Yeah. So we opened for him at Red Rock end of last year in Colorado. I don't know if you've been to the concert. I flew in for it. It was amazing. Just like amazing venue. [1:11:41] And we had multiple robots on stage just dancing, and they're all tuned into the music. And they kind of dance with it, and it was just, it was awesome. And we had them also here at our last holiday party in December. [1:11:52] And I don't know, we just raged with Deadmau5. Hey, it's Molly. If you enjoy our interviews, check out our newsletter, Sorcery.vc, where we deliver a once-a-week top deals and tech headlines email, and also go deeper on our podcast interviews. Subscribe to Sorcery today. [1:12:11] on YouTube, Spotify, Apple, or wherever you listen. Link in description to sign up.