I am blown away by how far you've come. The things that you can do with neural nets now just completely blow my mind. Every year to year, the whole business looks completely different.
It's amazing to me how you accumulate data and the data becomes this incredible barrier to entry, this incredible asset.
The one thing that's important here is that once one robot learns how to do a task, every robot fleet knows it. And humans still operate like this.
When do we start seeing robots building robots?
We will put robots on our Baky lines this year. Listen, this is going to be the largest economy in the world. It's going to be a super impactful business. It'll lead to ubiquitous goods and services for anybody in an age of abundance. And it's going to be a super fun business, too. It's going to build the sci-fi future we all want. What you're seeing is every major group in the world will get in the space. You have to. You have like no choice.
When are we going to see the first figure in the customer's home?
My best guess is I think...
Now that's a moonshot, ladies and gentlemen.
So Dave and I are in San Jose at Figure headquarters. We just did a podcast with our friend Brett Adcock. Extraordinary. And check it out. So we're here.
Yeah. Figure one. This is the original. Yeah. Still somewhat functional.
Yeah. It ran the first large language model, the first neural net.
They built it in under a year. Brett actually was screwing these things together himself and it was all about gathering telemetric data so they could build this.
Here's figure two. Much more beautiful, much more functional. Running neural nets across the board, dumping all the C++. Can it do live long and prosper?
Here we go with figure 3 is the workhorse right now. We just did a tour. I mean, probably, you know, saw a hundred of these walking through the hallways on test stands, cleaning dishes.
Was good. They added a flexible toe, too, so it can go down like this. And before it had this clunky clunky foot here.
And figure three has the palm camera.
Palm cam. Yeah.
They cut about 30 lbs off the weight and 90% of the manufacturing cost.
Wow. Crazy. Yeah. Amazing.
Yeah. It's the perfect height between the two of us.
Welcome to Moonshots, everybody. I'm here at Figure headquarters with Brett Adcock and DB2. Brett, it's been about 18 months since we did a podcast on Moonshots together. And I am blown away by how far you've come.
Months in AI time. That's like a decade, dude. Welcome to Figure Headquarters.
What do you think? Yeah, it's extraordinary. I mean just describe, we just went on a tour. You got 300,000 square feet, 400,000 square feet under development here. I mean there are figure three robots walking down the halls. There's fully autonomous robots I guess running Helix 2. You just released Helix 2 today.
I got it while I was flying up here. We have these robots doing everything from kitchen tasks to packages to manufacturing of different type. I mean how many robots do you think we saw?
Seriously, I wasn't counting as we... hundreds. Maybe not a thousand.
Yeah. Hundreds. At least 100 or so.
Yeah. Well, there's a lot of partial robots out there, too. It's hard to... picking up figureheads. How many hands do you think we saw? There's many more hands on the hand line, the headline, the torso line.
Actually, picking up the head was the most surreal.
This is where the pelvis is made. Yeah, for sure. Yes.
Pretty amazing. You know, I still remember during my first visit with you, you know, full disclosure, my venture fund is invested in two of your earlier rounds. Super proud of the progress you've made. I still remember your figure one putting a Keurig cup in a coffee maker and that was a big deal because it was done with neural nets and not C++.
I mean that honestly was like I think that was a big inflection point for us. I feel like the, you know, I think a few things we needed to really run down is can you build an electric humanoid? It's like low cost. It's capable like a human, just the hardware side of things. The second thing is can you figure out a way to not code your way out of this problem? How do we use a neural net to learn those like human type representations and the new task? And when we were doing the Keurig task, it was basically a bimanual neural net running on the robot which is now like evolved into Helix. And I was able to basically do the whole kind of like, you know, it was a smaller task but was a few minutes longer of like picking up the Keurig cup, opening the coffee, putting it in, running it, and it was the first time we saw like true instance of kind of like neural nets working on a bimanual humanoid robot.
Yeah. And that was when we were like, 'Okay, we have to just go all in on neural nets.' The whole stack needs to be neural nets to make this work.
And that started like, and that was basically 2 years ago now. And then you guys saw Helix 2 today.
Which is like the basically like the best release we've ever had. So, we'll run a clip of Helix 2 while we're describing it, cuz what we saw was figure 3 running Helix 2 in full autonomy, going into the dishwasher, picking stuff up, putting it away, not pre-programmed.
And I loved the human elements of it, like using its hip to close something and its foot to raise the dishwasher.
That's the neural net difference, though. You get unexpected behavior, you know, both good and bad, but things you could never code up.
You could never... like your career went the software company, VTOL company. Now, this has got to be the first neural net platform.
Yeah. The things that you can do with neural nets now just completely blow my mind. Versus code, like we could have never have done a quarter of the stuff that you saw today with the whole body, with manipulation, with things that you like, you know, there's only so far you can really push like code heuristics into a humanoid robot, it's just a dead end. It's just not going to work.
Yeah. It's amazing to me how you accumulate data and the data becomes this incredible barrier to entry, this incredible asset.
If you were writing all this in C++ code, that code would be, you'd have millions, hundreds of millions of dollars invested. You would not want to mess it up. With the neural net, you can say, 'Look, hey guys, retrain it from scratch right off the run.' It's just a completely different approach. And that's why people are way underpredicting how important or how quickly this is going to evolve cuz it's a completely different paradigm.
Well, we've like lived through it. I mean, like we, you know, I think maybe a year or two ago, we had like several hundred thousand lines of C++ code.
Several hundred thousand.
Yeah. Probably 100 bucks a line to write it.
Yeah. Very expensive. Very hard to like test and like get out reliably. And like also hard to like model all the different behaviors that we would need to test the Keurig. And then you know we removed a majority of all that in the Helix one where we still had a lot of like lower body control being run in basically the control stack in C++. And then today we removed the remaining 109,000 lines of C++. All neural nets today. That's a full body and that took it from like being able to do really good tabletop manipulation like you saw the Keurig coffee, the work we do in logistics, all that being done in neural nets, we've been showing like amazing progress there, but getting the whole body to get out of there and move dynamically through a scene while manipulating and planning is just a whole other, we basically spent like a greater part of a year refactoring the Helix architecture to be able to enable this to work.
You're talking about now like moving through space like a human.
Yeah, having like control of the full body, eye, hand, foot, leg coordination, everything is sensor data in cameras, tactile, we have palm cameras, basically doing inference on board the robot fully embedded and then be able to output torques into the motors and do that, you know, a few hundred hertz, you know, in terms of like, you know, that planning control and do that reliably on very difficult tasks. These are bimanual tasks where it's grabbing and holding things, planning, moving the body, getting things out of the way. Making like errors and replanning and fixing this. All done with the neural net now end to end over like a pretty long, for us it's like, you know, it's kind of like room scale autonomy. So we can like now like finish the whole room which is important. And next we're going to graduate to like basically the full house.
That's one of the things really obvious when you're walking around looking at what everybody's doing and working on. You visualize a robot company having lots of people working on microcode or actuators or batteries or whatever, but there's just a huge number of people out there at workstations. They must be working on the neural nets cuz you know it's just got to be such a dominant part of what makes the thing actually look and feel human and you know the motions are so smooth and you know everybody when they think about the history of robotics they kind of chart these line charts but it's not like that. It's a disruptive change from dropping that last 100,000, 105,000 lines of C code to moving to an all self-organizing neural approach.
Completely different future. It is like we make these like technology progress steps and I think it's been very apparent here like every year to year the whole business looks completely different.
Yeah. In large part of trying to get the hardware, hands, like all this stuff in a good spot and then be able to have like more range of motion and speed and torques like a human and then be able to get like, you know, we're all in on neural net so it's been like, you know, what is the right data set for that for pre-training and post-training, do we have the right training cluster, do we have the right models, and then deploying those really well on the same humanoid hardware that's like a full loop like we've actually designed figure 3 to run like, if you say like what is the guiding principle of figure 3 more than anything else, it was just designing for Helix. How do we design this to run like Helix on? How do we get Helix a bot?
So counterintuitive.
Everything just, so build around the neural net. We looked at the neural net and we said how do we like fit like this into a humanoid robot and what are the best sensors? How should it run? What does the operating system look like? Middleware, firmware, embedded software like all of it is encapsulated in this like view that we need to go all in on neural nets and do humanlike work.
About to release the 2026 version of my humanoids metatrend report. It's a deep dive looking at a 100 different robots in development right now. A deep dive into 10 of them including figure, 150 pages. You can check it out at Substack for my paid subscribers. Anyway, super pumped. This is a field that's moving at exponential, hyper-exponential speeds. So in the beginning you had partnered actually with OpenAI on software and you made a departure from OpenAI and I mean I guess are...
Not quite accurate but like...
Okay well you can, you know...
I mean I think, you know, I met Sam and the OpenAI team and they were just really interested in getting into robotics and it was like in their early master plan is to get into like basically shipping home robots and they really wanted to kind of like basically work on a very intimate relationship. They ended up leading our, co-leading our series B along with Microsoft and we started working on basically like a collaboration agreement to help like work on next generation models for humanoids and we were like super big then still are on like how do we like language condition the whole stack, how do we use like an LLM in a lot of ways is just like this like world model. It really understands like in the weights like basically like what things are, what it should do, has a lot of like good semantic understanding.
We're trying to how do we tap that in for the humanoid? How does it, how do we learn from this at scale and some of those representations. And it just like the partnership just didn't work. Like my, our team just ran circles around them for basically better part of a year and it just came to a point where like it just made sense to just, we were just doing all the work ourselves internally. We had a whole team here, a lot from like some of the best like labs in the world and we were putting out like work after work. The Keurig coffee stuff was done by us. All the stuff was done internally and at some point it just didn't make sense to train other folks on how we basically build AI models internally for embedded systems like a humanoid.
Did it turn out that that LLM matters at all in physical... like you could start with an open source LLM and...
Is it like a vision language action model you're...
Yeah, we basically like, I think the LLM is definitely a certain piece of this. Like we basically want to like take it like the semantic grounding like a VLM.
Yeah. Like the common sense, you know.
Yeah. Like how do we like understand from this? So you know, which we have in the Helix today, super critical but like getting to a point where we can understand physics in the robot and have it like really be able to plan and reason at fast dynamic speeds was something that nobody in the world's ever really done before and I think that that's the work that we I think have been excelling at and really love is just like how do we get us like understand physics.
I think most of our audience probably knows this, but just to rewind the tape, you know, the LLM, GPT-2, GPT-3 built entirely on text data, scraped right off the internet.
And then they supplemented that with a ton of other data also in text form. And that creates this machine that has tremendous amounts of common sense. And if you ask it, hey, do you know how to play soccer? Says, yeah, of course I do. But then you try and install it in an actual physical moving machine, it has no idea what it's actually doing.
Yeah. I mean like we basically need to like touch everything in the world and we have this like really high dimensional robot that has like, you know, 40 plus degrees of freedom. So like, and so like, you know, on the surface area just the math around this is like the dimensionality space is really high. So you have like 40 motors they all can spin 360 degrees. So the amount of states the robot can be in like positions is like 360 to the 40. So there's more states of the humanoid than atoms in a universe.
So you're not going to simulate those one by one.
So like, so the question is like then I need to like understand these fine contact dynamics of like I need to grab this water bottle like where do I position my elbow, pelvis, like torso, head like fingertips. How do I plan the, you know, to grab this and then how do I put pressures on there and how do I understand those representations really well from observations now into actions that I'm doing at test time. And this is not in like, knows none of this, knows this is a water bottle. It probably knows like I need to grab it from the side. And then but like all this implied physics that we need to do here just we have to go train models to go do that.
It's actually kind of weird because it thinks it knows how to do it too. You know the LLMs feel like they can do things, you know, intuitively and then they completely fail.
I mean you can, we've done this, you can zero shot the LLMs inside a robot. We do it, we still do it actively. They just can't do anything.
Just for fun just to watch them fall.
Yeah. I'm like, you know, I'm kind of interested. Like a project I've been doing is like, can we just like, the other day I was like, can I zero shot? Like I'm working on this new AI lab that I founded recently called HARK. And we have this new AI model here that is like it's just completely...
Wait, you founded a new AI? Well, rewind the tape here. What?
A new AI lab. I sent you this.
Yeah. Working on and we have like some new AI models and we actually put one of them into the figure robot like this month. And I was like, okay, let's just zero shot. Let's give the LLM, you know, let's give the model, this is like a multimodal model. Let's give this access to just like basic commands like basic like XY coordinates, like basically can we give it like acceleration and XY coordinates for navigation like basically a joystick. Can we give it a digital joystick?
And I asked it to like find the exit sign and just like go to the get out of the building and it just, and unfortunately it was going the right direction and ran into like a clear glass wall.
Exactly. So, we've really stressed this. It just doesn't work. Like you're missing so much world understanding of like what's really happening? How do we move my body? Like we're thinking like, you know, it's pretty simple to grab an object maybe with a stationary robot, but the robot we have for humans are moving the pelvis and head and torso and hands and arms. Like when you're reaching out to grab something over a table, your pelvis is moving backwards. Like it's very difficult to command like very high...
Robotic physiology. You know, out of China this year some of the government employees said we've got a robot bubble. They have 100, I don't know if you saw that article that came out, we have 150 plus robot companies in China and I mean there's a lot going on there. You know, in the US I would say maybe there's 10 serious players, I mean two or three who are extremely serious including figure. There's but there's a lot of potential humanoid robot companies. I was just at CES and saw, you know, I mean it was a humanoid robot explosion.
And then as many or more hand companies which is interesting. So I go back to sort of the early 1900s when there were like 250 car companies and like two or three hundred tire companies and then this massive consolidation occurs and GM and Chrysler and Ford sort of buy and consolidate. What do you think is going to happen with all the robot companies today?
Um I think it happens in every industry like this especially in deep tech. This will all consolidate down to a few groups globally.
Do you have a guess? Is it a triopoly? That's the right description. Is it, you know, more than 10, less than 10? It always seems...
Globally. It always seems in the US anyway to settle down to two, three or four. But the borders are not obvious. Like with cars, cars are cars, right? And actually you had cars and trucks and those were kind of separate for a while. Well, you also have different designs like I want the plush interior. I want the sportster. I mean, and that I think I wonder is it going to be, are robots going to be differentiated by their vertical application, their personality?
Exactly. There's so much more variety possible in robotics.
Yeah. I think everybody's just like taking for granted how difficult this is.
Yeah. Like this is like you have to go out and build like basically like pretty novel, very very difficult hardware. Needs to be relatively cheap. Then you got to figure out how to make neural nets work on it. And then you got to make neural networks work on it at scale. And then you got to manufacture at scale. And then you got to get these products out reliably that all work every day without any human intervention.
And you know I think we talked a lot about this like how we're doing like Keurig coffee work. Like I haven't seen any single human in the world do that or able to do that today globally and that's been two years.
Yeah. I mean by the way a lot of the video we see is actually teleoperation. I think I wonder if people realize that a lot of the robot companies are teleoperated versus fully autonomous. What we saw just walking around here was a 4-minute long fully autonomous operation on Helix 2, right?
I've never like, I've built a lot of businesses in my day. I've never seen so many companies with a human in the back commanding the robot and putting out updates in my life. I've just never seen it. Like when I first started thinking it was stuff was coming out, but now it's like every week is somebody just telling a robot and putting out a video and it's just...
It'd be the equivalent of like I have a self-driving car company and there's a guy in Tennessee driving it and we're like marketing as like there's no humans in it, self-driving. We're putting out teasers. We're in a lot of cases now there's companies selling a service like, so I think like, I mean if you want to do this right you got to believe in neural nets all the way down the stack you got to basically build for general purposeness.
So the parameters that make, that's going to define the successful top two, three, four, the neural nets, manufacturing...
Okay I would say like I would say what's impressive today is not manufacturing, you could probably solve. You know, we're pushing on manufacturing hard, but you can probably solve general robotics with 100 robots. You impress like a full end to end robot that is generalizing to an unseen place. Like you can drop it into an Airbnb and be able to do long horizon work with neural nets.
I mean, any long horizon work in unseen places.
What do you define as long horizon? Hours, days?
I would like to see days of work. Yeah. Full time as days of work and at least at the very least. And we're like so far from that.
You have like robots out there doing like karate and jumping, which is like these are like pre-programmed open loop behaviors. They're not impressive. We do that. We've done that stuff here. Like, you know what I mean? We've done like the open loop behaviors. It's just there's just like any college kid in a dorm room can do this with a robot and...
Yeah. So I think like that plus teleoperation, teleoperation is not impressive. You could build a shitty hardware and still teleoperate it and put out videos. That is not hard.
What's hard is to do full end to end neural network in unseen places or generalize to this. And then if you can solve that then the next step is like how do you get that out at scale? But we are still in the like who can solve general robotics phase of the humanoids phase. And it's just not impressive if I can build 100,000 robots right now that like need teleoperation or can just only do open loop replay. Like it's just not cool. Like if we're like, Brett, your only job is to build 100,000 robots right now.
We have the capital to do it and we can do it. But like what we really want to solve is like I can give you 10 robots and they can go into unseen places and do real useful work. Like that's what's going to differentiate.
So iterate that until it's right and then mass produce.
Yeah, you basically want to be bringing up mass production in parallel cuz like building like high rate manufacturing for humanoids is going to be super hard and you're going to like have to go through like a lot of iterative design process. So it's what we're doing now. We're bringing up higher volume manufacturing as we're like learning how to build true general purposeness. So but my view is like if you think about these like level bosses that happen that will hurt, like you need to graduate to doing like, you know, first very short periods of like neural network which we haven't seen a lot of in the world today. I don't think there's anything over a minute long in the world that's doing neural nets continuously today in humanoids.
Everything's cut, all the films are cut or teleoperated, it's pretty crazy.
I'm really glad you're telling us that.
Yeah I mean like that's, you watch any video you want, you want to see it uncut, you want to see done with neural nets, it's like not teleoperated, and then you want to see stuff we showed you here in person today that are running for hours and hours and just like we run these robots with no...
The kung fu videos whether they're teleoperated or fully autonomous are actually fascinating and scary when you see them.
But the technology around there is not great. I mean you're basically putting somebody in a mocap suit, you're having some guy like do karate chops or walking around and then you're running that open loop, you mean you're running that blind, you're just hitting a replay button.
And you can do that with a very simple like RL neural net. You can basically do mimic on this and it's super simple like there's like open source code for this you can do with like basically one GPU on your desktop and you can do with any robot and every robot has a very tiny amount of computer. These are like single million parameter models, are very small, you don't need a lot of memory and they're very simple to execute. What you really want is good closed loop control where it's reasoning like over like kind of like 200 hertz or 200 times a second. And it's dynamically responding to the scene.
And that is literally a million times, 100,000 times harder than doing open loop.
What the human sort of cycle time is...
I mean it hurts. Oh gosh. Much lower than that.
One thing we've seen about a robot is we can like balance on one leg like better than a human.
We just have like much better faster dynamics.
Can we talk about the speed of development here? So 2025, I'm just trying to imagine you put out this beautiful post every week on X about the progress in the robotics field and what's going on here at Figure and it just constantly, you know, locomotion was a big step forward, excuse the pun, for Figure just seeing it walk and then run very naturally. What else was significant in 2025 for you?
I mean we launched Helix in 2025 about this time last year, about a year in now. I think that was like highly significant like we basically figured out how to run like basically like longer, like over long periods of time neural networks on a robot. How do we get the data for it? How do we train models? How do we deploy to test time? How do we get like, you watch like package logistics? I think I saw it running there. It's like running for...
It's been running for days now.
And it's just like it's a neural net all the way down the stack. It's learning how to grab packages, kind of like, you know, individualize them, find the barcode, position it down. It'll even pat the package down so the barcode reader below can see it and scan it. And it's doing that on very high accuracy and assume that high speed.
High speed though. That's the part that... Yeah. No, it's visually crazy.
Yeah. Because a lot of what you see in robotics is it's as fast as a human would do. Yeah, we like our last, like we see air now, like last room we did, we had one air over 67 hours continuous.
This thing is crazy. It's doing an operation every second or two so 67 consecutive hours of that is a lot if you had to guess at...
So I would say Helix is a big one and then figure 3. Figure 3 is a huge step change for us in hardware.
If I could, what do you see then going in 2026? Here we got the, you know, next 11 and a half months, what are you excited about?
We will build like our entire road map around Helix 2. Now we will basically, now Helix 2 can like go from like doing the logistics use case stationary to walking and moving and basically do like long horizon full body control. So that means the, and we basically have now integrated all the sensors, tactile, camera, palm into the stack and we're seeing like improvements overall in the policy layer. So we're getting better and faster about like basically like taking data and basically running it on board the robot now.
So I wanted to ask you like what defines Helix 2 because you're probably incrementally improving the neural net every day. Yeah. So what is the...
A couple big steps? One is we basically have an integrated basically a fully learned what we call like system zero which is our controller into the robot. So the robot has a full body reinforcement learned controller in it.
So basically now we have like literally no code running on that robot. So it can basically move the whole body itself using a full like basically like learned controller inside of Helix. We call it S0.
Has anyone else ever done that before? That's got to be...
There are reinforcement learned controllers out there. Like a lot of the karate stuff you see and things like that, but nobody's really integrated that in the whole body for learned manipulation and perception. And nobody's showed that actually working with like moving around and doing things that we saw today. I actually don't even know if anybody showed it stationary standing and doing learned policies. Actually, probably not in the world. So like getting it integrated into a stack now that we actually use going forward. I think one of the things we learned at, like we were in BMW last year and we were there for like we did six months like we deployed like our figure two robots every single day.
The biggest thing we learned there is like the stack we had I think about 80% of the things we got right and 20% of the things we got wrong. Meaning like the things that we got wrong on we didn't want to scale. It was working. The robot ran every single workday and it worked. But then we learned like okay I don't want to ship 100,000 robots in this like architecture stack. It's just like too hard to scale. It'd be like too brute force. And so we basically worked on basically for almost basically a year now on like okay what is the ideal architecture where we can go out and accumulate large sets of pre-training data. Put in the robot and it can just like do this work and we emerge generalization from this and that's what you're seeing today.
Helix one had the C code in it still. So what defines Helix 2? Helix one had a lower body controller...
That was still written in C++ and everything else full upper body was full neural nets.
And so we basically completed now the full body.
And then in doing so we also like did some work on our system level, system one level where we integrated all the sensor modalities now from the hands and the rest of the robot into the stack. So like for example, we now have tactile sensors in every fingertip that we're using on Figure 3 as well as palm cameras to understand how we're like, we're sometimes occluded and sometimes we want to better basically better understand how we're grasping items. So we put a bunch of stuff out about we're picking pills and stuff out of pill cartridges that you like literally occluded from the hand. Your hand's like literally in front of the head camera, but we still really want to understand where we're going.
So I think, so now with figure, with Helix 2 we basically have a full stack end to end with neural nets and we feel confident scaling the pre-training data set into Helix. I even go as far as like we've designed Helix 2 for the pre-training data set and then we designed the robot for Helix 2. So we've designed everything around data.
And how do we get data at scale? If you're in the neural net game, it's like a data play. It's like how high quality and diverse.
So it's experience. It's just gathered experience in the field in all kinds of circumstances.
It's weird. And I know everybody knows this already, but it's accumulating and it never goes away. It's incredible progress.
Well, yeah. You teach somebody how to scuba dive or how to play piano and they have that knowledge. They live, then they die, then you have to teach somebody else. This is completely accumulating.
The reason why I think it'll be like a very few human groups is like the one thing that's important here is that once one robot learns how to do a task. Every robot in the fleet knows it and humans don't operate like this. I wish we did.
I watch my kids like I like kids like learn how to do stuff and they just don't listen.
So 2026 predictions. What's your boldest predictions for figure? What is your goals for this year? What do you imagine?
Yeah, I mean we basically we're spinning up Baky like production enormously right now for Figure 3.
So you said something like a robot every 30 minutes you expect?
We're trying to get there in the near term right now.
Which you guys saw we walked through Baky today. What did you guys think of Baky?
I wish everyone could see it. I guess it's all secret. You can't go through there. We haven't, there's a lot of IP there cuz like you see exposed boards and actuators and stuff.
Maybe we can mix it in here. But so whenever there's a lot of humans, when do we start seeing robots building robots?
We will put robots on our Baky lines this year.
And then phasing like basically like phasing humans out of there will be a combination of getting more robots there and doing more high volume like automation over in Baky.
Okay. So that's the first 2026 objective.
We want to scale up robots in Baky for sure. The second thing is we want to scale out robots in industrial commercial workforce. So we have like multiple clients now we've signed, they are like buying or leasing robots from us and we are going to get those out at scale in 2026. We have we know exactly where we're going geography wise, what use cases are going to be, deployment schedules. We want those to be figure 3s. So we've just retired end of last year figure 2s and now we're basically building the arsenal of figure 3s out as we're scaled up manufacturing to get them out into the world and run every day. We like the commercial workforce because it really helps harden our ability to run robots every day. Like what we're here to do is like we're here to build robots and run them in the world and they run 24/7.
Your ideal customers who I know a lot of people would love.
We have to be frank like so much demand for customers. We have like we've talked to 50, 100 customers or so in the last like 6 to 12 months. We really want to be like kind of all in with a smaller group of customers and really spend time with them, integrate well into their facilities and you know, do well. We're still at this like, we're still early right. We don't have like thousands of robots right at these places. We want to as fast as we possibly can, but like once we get to this certain like, I mean we could probably ship like I think we could ship an enormous amount of robots into the current customer base we have now. Like so we see like, you know, we're kind of good now for the next like two or three years in terms of like we have so much demand like they're kind of waiting for us to like ship at scale.
Yeah. We have like a service, we have like a, you know, we really like the leasing model.
So you know, thought of it that way, used to be a bot at least these days.
Yeah. You lease humans. So we like, we lease humanoids today. You know we won't be like, we're not opposed. I think what really matters is trying to figure out how to find the right distribution to get robots out at scale. Like it'll really help us get really good at what we do. Like it's one thing to like show a demo or whatever else, but like, you know, when we had robots at, you know, in a commercial customer last year at BMW like it was just, it taught us a ton about like running it every day, fleet operations, safety, like repair and maintenance, like there's a lot of other things that need to come through on the ecosystem that we need to get right. So I say second thing is like getting robots out at scale commercial customers and then the last thing which is arguably the most important for us is we want to solve general robotics.
Yeah. We want to basically like the analogy is like we want to build a human in a bodysuit that you can just talk to that has like common sense reasoning, you can communicate with that has like almost perfect memory what's really happening or what's going on in your life that can maybe talk to you, almost be your companion and then go off and do things that you would like, like an everyday human would want to do and I would expect them to get up to speed on those tasks at or faster than a human can.
Is there two different models then driving it? The VLM model for the body and the physics and the embodiment versus an LLM for conversation and memory.
We believe this all comes down to like one model at the end of the day that is one omni model that is trained early in pre-training that helps fuse all this together. But yeah, you could think of it like we need to have speech, we need to have like language conditioned policies, we need to understand physics really well. We need to remember things and be able to recall that easily. We need to have some sort of personality on the robot. I think one thing that you're going to see more and more is we really want to make this robot something you can spend time with.
And we've been really focused on getting the core building blocks built. But like over the next year or two, I think you'll see us, I think I just want a robot in my home I can talk to like remember things. Talk to my kids. My kids come home like sad from school or something. I want the robot to understand that have the EQ like self-awareness to see that talk to them. Like I think all this is like something we want to spend more, we're spending more time on now internally.
Is there already a big model where it'll have different like depending on the task you're doing it'll run different parts of the neural net or does it...
We have like one neural net now that's like, that's basically there's no like libraries of neural net that we pull down.
That's interesting. So there's no like a dishes neural net or like logistics neural net...
Yeah. Because at scale, if you teach the thing every physical motion, massive number of combinations, the storage is actually dirt cheap, but the processing is very expensive.
Yeah. Even better, we've basically seen positive transfer now with all this data. Like coming in, the robot can generalize better with more information. Even like more knowledge is better. It does cross, like playing piano makes you a slightly better soccer player. But also, you don't want to run the whole parameter set for piano playing when you're playing soccer. It's an interesting little hybrid problem there.
Yeah. You don't want to nuke it. Yeah, for sure.
Yeah. I mean, that's where we try to build best-in-the-world models here and build a great team that can ultimately deploy robots that are useful. I think showing like, you know, this type of usefulness, a lot of the stuff you saw today, and a diversity that is super important for a humanoid robot needs to be able to do everything a human can, which is like the distribution curve. It's like, you know, we probably do like billions or trillions of unique things in the world.
Yeah. One of the things you said on our tour that totally tells me we're on the right track is that you're using normal GPUs for the training like everybody, but the inference time compute is on super, super fast dedicated non-H100, non-GB300 hardware.
Yeah. It's also running fully on board.
And it's running fully on board. So we can basically do very fast inference and policy deployment.
And it's also not sucking down the entire power of the robot.
Yeah. I mean you also have an issue where we've also run models offboard the robot, but if we lose communications or have some...
Yeah. You know what I mean, like if you lose internet, it's hard to do work. And mostly we hit on supply chain, batteries, and comms. So on the comm side, do you imagine we're going to be running a 6G network on there besides Wi-Fi? What's going on in batteries these days?
We have, so from a network perspective or a communication back to the robot, we have Wi-Fi on board. We have a 5G and SIM card, eSIM on board. So the robot, you can text the robot, it can have a network outside of a Wi-Fi condition, and then we also have Bluetooth on board. So almost like a walking phone or something like that you would think of.
So we want, I mean, you really want connection at all times, but you also want, ideally you want connection all the time, but you also want the robot to be able to perform work without a connection. So you really want a lot of onboard intelligence, you know, that we basically, in case you lose internet, the robot's not bricked. I mean, humans for the most part can do work without their cell phone.
Not teenagers, that's going away.
Not teenagers. Okay. So, batteries, I mean, they've been improving. What's the battery life right now? I love the charging mechanism, by the way, for those who don't know. You're charging basically through your feet.
Through your feet. No connector. You just stand and charge.
It's really cool. What kind of battery life are you getting? What do you expect in two, three years to get for battery life?
Yeah, we run basically around four to five hours per full charge on the battery. And if we're starting at full battery life, and then through full depth of discharge, and then we can charge wirelessly about 2 kilowatts through the feet inductively. So we have about a 2 kilowatt-hour battery pack. So it's about an hour or so for a full charge on the robot. So we can do four or five hours on, an hour off.
And yeah, it's great. I think folks are over-indexing too much on how long the robot can run on a single charge.
Yeah, I don't expect there are that many tasks that take a few hours.
You're not, you know, go take a little break, do this stuff. So I think there's ample time to do opportunistic charging. Maybe send another robot in. We also can charge, we basically can put this little thin mat anywhere in the world, like it could be a conveyor system or wherever else, could be at home in front of the kitchen. You can just charge there while doing work, which is really cool.
So you don't have to have any wires or things like that. You're pulling from the...
Well, I think one of the greatest value-adds you're doing right now is people are over-indexing on all kinds of weird things because they're physical beings and they're watching the robot do physical things and they're saying, 'Oh my god, can you believe it can sprint now? Oh my god, it can do a backflip now.' And you're like, well, it depends whether you programmed that in C or you tele-operated it or did it actually learn this?
Yeah, I think most of those are open loop. They're just like replay buttons.
Yeah, exactly. And it's just so hard. So when people say, 'Well, how long does it run with one charge on the battery?' you're kind of relating it to your cell phone, but it's not relevant in the inflection we're going.
Yeah. I think you just got to, the summary here is just like I need to see real closed-loop control of a robot moving around, touching and moving things like a human would.
And that's where the hardest problems all sit and that's where we've seen this huge wave of humanoid explosion, like I think you said out of China. But we've seen this very steep drop-off from getting to that point next, which is even like show me a minute of the robot doing Keurig or something like that, uncut, closed-loop, real-time.
Yeah. And I just, you just haven't seen that.
And I think you will and I think there's a lot more levels to go from there. And that took us two years to go from a few minutes of tabletop manipulation with neural nets to a point where we can do kitchen work, room-scale autonomy. And that was two years of working seven days a week, a lot of nights getting there. So it just gives you a little sense of like you're not going to do that in six months from there. So I think there was a lot of both hardware, low-level software, firmware, embedded systems, sensors, and then neural nets and then data. All of that came together to build this. We couldn't have done this work on, we couldn't do the same work today on a robot that we could go buy today off the shelf.
You vertically integrated. I made the choice to vertically integrate, but supply chain, how much supply chain ties back to China?
I think by summer we'll have almost none of our supply chain in China anymore.
And do you buy into the US versus China sort of AI and robot competition? How do you think about that?
I don't, I just, I spend a decent amount of time in China. I love it. China is great. Like I go there and it's, I know you're watching TV here in the US and it's just like this massive conflict and battle and everything, and then you go to China and everybody's just trying to help and win and trying to work and collaborate. It feels like a startup incubator and it's like a fun, one-way competition. It just feels like everybody's on team human, team humanity to go win and it's so great. Then you come back here, you're poisoned with all this stuff online and articles and television and it's just not like that when you're on boots on the ground. It's like let's go as one and go win. And I just love that spirit of trying to just progress this technology as a giant lever arm for humanity to bring abundance basically for everybody and just make it a sci-fi future we all want to live in.
Oh my god, that is, we want to speedrun Star Trek is what we talk about.
Exactly. It's like I see Figure on the moon, Figure in orbit, Figure on the ocean floor.
100%. So, the equivalent, you guys make your own actuators, motors here, and part of that is because you want the exponential growth effect, but part of that also is the supply chain just doesn't exist to give you the parts here.
Improvements, I mean, you're talking about the improvements made between Figure 2 and Figure 3.
Because you have all of the ability to iterate in terms of speed and cost. I mean, the numbers that you shared on cost, I was like a 90% reduction.
Yeah, we reduced cost like crazy on Figure 3.
I think we, listen, we vertically integrate because we have to. It would be great if we can go off and buy motors and plop them in the robot. It doesn't work like that. It'd be great if we can go buy hands and just screw them on to the end. It just literally doesn't work. If you go through the engineering work to basically understand how we do comms and power and sensors and failure cases and thermals and low-level firmware and embedded software, it just, there's one of the cost, something breaks, and reliability, something breaks in that equation and you're left with hopefully the vendor fixes it or you die. It just doesn't work. None of the stuff, the technology readiness of these things are really low.
We would have loved to have gone out and bought all of the stuff in the early days. We tried and we basically just failed at all of it. So we had to go design ourselves. And then now we manufacture, we do all final assembly and everything here. And we do that, you know, in some cases because nobody knows how to do that well. We do that a little bit for IP, like we really want to control that here and understand what we have people have access to. And then we also want to get good at making a lot of robots. Like what we need to get good at long-term is probably a few things: getting data at scale that can run neural nets, and then doing Helix really well, and then making a lot of robots, and then getting those things out at the world at scale. Like a pretty simple equation during the day.
It just feels like the journey of getting Figure up and running must have been so much harder than it would have been in China. But then once you have everything built in-house, all the actuators, training the neural net and everything in-house, then you have a massive advantage versus anything going on in China. Because if you'd been locked into a supply chain, it only has certain models.
It's like even if we used an existing supply chain for all this stuff, the robot wouldn't be able to do what you saw today. It just can't do it. If you go out and buy a humanoid robot off the shelf today, we can't get it to do this.
We've bought robots off the shelf. We've looked at them, like you just can't get them to do this work. They don't have the right sensors. They don't have compute. They don't have thermals. They don't have the right hardware, the hands, the head. All of these are built around a neural net stack.
Yeah. Well, that's the new thing, too. The neural net is incredibly integrated with a specific hardware.
If you watch folks that are trying to buy these robots off the shelf, like say from China, they'll end up retrofitting them themselves with these giant backpacks. They'll have power, they'll have compute there, they'll have thermals, they'll have a wire hanging out. They got to hook that into the back. Probably has its own local battery. They'll hook it in the back of the robot. They have to take it and they have to overclock it and it's just the wrong way of doing this.
It's like buying a rocket and putting stage two on the side or something like that. It just doesn't really work at scale. It works for hobby-grade demonstrations and things like this, but if you really want to do robotics at scale, you're going to have to go design yourself.
Yeah. Looking at the companies coming out of China, Unitree, Engine AI, and so forth. Do you have any which are the ones that you're most interested in, excited, as friendly competition, if you would?
Yeah, I think one thing that's great about China is there seems like, as you mentioned earlier, this explosion of really great talent and robots coming out the door.
And great entrepreneur work ethic there, right?
It's awesome. Like, it's great. And I think it's just good for humanity and this needs to happen. I think the thing that we've not seen is we've not seen any closed-loop AI control from these systems at all.
We've seen a huge lack thereof of that stuff. I mean, industry is like, here's the robots, we'll sell them, and they're doing a ton of basically open-loop, local...
They're hand controllers.
Yeah. So I think doing that is very, it's almost orthogonal work from designing the system the right way for full autonomy. But there's, I think, if we think about who Figure really competes with as our main competition, it's certainly China as a whole.
For manufacturing, I mean, for human labor.
I think just for humanoids, we really don't see anybody else besides China as a real competitive threat today.
Fascinating. Do you, rumors about Apple getting into the business? They cut down their car project and the rumors are that they're heading towards humanoids. Have you heard that?
We've heard this. We've had every major, I've been in conversations with every major tech company in the world last 12 months.
And then Nvidia and Google and even Sam. I mean, everybody's making noises about...
Listen, this is going to be the largest economy in the world. It's like half a GDP, roughly a little under half a GDP of human labor.
Yeah. This is the next great place to be. And I think it's going to be a super impactful business. It'll lead to ubiquitous goods and services for anybody in an age of abundance. And it's going to be a super fun business, too. It's going to build a sci-fi future we all want. It's going to feel like 2080 up in here. So what you're seeing is every major group in the world will get in this space. You have to, you have no choice. You have to.
Major group being Apple, Microsoft, Google.
Every, I think every major player that wants to do this. I think the thing that, I think it's going to be hard is we're doing rocket-type difficulty in design here.
So, it's like if Meta is building rockets, you'd be like, that'd be crazy.
Yeah. And I would think maybe the humanoid is probably up there with rocket design. It's certainly harder from an engineering perspective than when I built Archer, building electric aircraft, and that was hard.
Yeah, that was a very, these are 6,000-pound aircraft, 12 motors, six independent battery systems. We built our own control stack and embedded systems. It's a very, did all the structural design ourselves, things like this. So I think it's probably up there with some of the hardest hardware on the planet and you just have got to be all-in.
Let me ask you about that because we were talking backstage at Abundance 360 last year and you had the basic tech stack had six layers of competency. You could probably rattle them off the top of your head actually, but yeah.
This is for our turf, for robotics prior to neural nets, I guess. So it applied to Archer and Figure. Sure.
But what were they again? It was...
I mean, Archer was basically a flying aircraft. So I basically built electric vertical takeoff and landing aircraft, right? That is basically, sorry, it's like a flying robot is what it meant. That has battery systems on board, has electric motors. Electric motor is just basically a stator, rotor, gearbox. Pretty simple. We have a little bit more sensors in our actuators than that, but for the most part, there's encoders and stuff in there and things like that. You have basically control software, like how do we control this thing and make it move around. And in the case of Archer and Figure, it's a very over-actuated system. So Archer has like 24 degrees of freedom. We have propellers, tilting, we have pitch on the blades, we have flaps on both the tail and the wing. And then, you know, Figure, we have over 40 or so on the system.
You have embedded software on board and sensors. So how do you get the compute, sensors, and embedded software all to talk to each other?
Then you have structures.
So those are like kind of the core ingredients of a robot or something physically moving through the world.
Yeah. So then my question is, traditionally the employee base would be experts in one, two, three, four, five, and six.
And they'd be really, really good.
So then you come in and overlay this with Helix and you've got this massive neural network thing. Is that a seventh competency or is that something that permeates the other? Or did you take all of your microcontroller experts and start training them on neural networks? Like the next thing, Archer, is how are you going to plan? And you do it through a pilot. We have my aircraft, Midnight, is a piloted passenger aircraft. So who's doing the planning and higher-level behaviors in the stack? It's all the lower-level control and code what to do.
And here at Figure, it's been changing over time, but now it's entirely neural nets with Helix 2.
So it's like who's going to, I mean, what is the highest-level behavior telling the rest of the stack what to go do?
Yeah. Where's that come from? A human, it can come from a joystick. It can come from an open-loop behavior which we talked about before, or it can come from a neural net that's doing the planning and reasoning. So like the kitchen demonstration you guys saw today and we released, what's telling the robot what to go do next and to pull the rack out of the dishes and to go grab the cups and not the coffee cups but the water cups. That's a neural net making the planning.
In the case of my aircraft at Archer, it's a pilot determining when to take off, when to hover, and when to transition into full flight.
And then how to sort of descend, different type of neural net.
It's a human biological neural net. Yeah.
This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Engineers start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and pre-compiles code for each task. Blitzy delivers 80% or more of the development work autonomously while providing a guide for the final 20% of human development work required to complete the sprint. Enterprises are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their org. Ready to 5x your engineering velocity? Visit blitzy.com to schedule a demo and start building with Blitzy today.
Let's talk into application layers. So, we're seeing your movement into the home besides the industrial base and such, and healthcare is going to be a big part of this. Elder care, helping people stay healthy at home. By the way, you just came through Fountain Life.
How was the experience for you? Thanks for referring me.
Yeah. It was great. I went down to a clinic a couple weeks ago.
Orlando. Yeah. Headquarters.
I didn't know what to expect. You know, I've done like full-body MRIs and CT scans and blood work before, but like, got there, it was basically a full stack. I mean, you know this, but it's a full stack.
Everything measurable about you.
Yeah. Exactly. Like 200 gigabytes of data.
Exactly. And spent like 5 hours there, left, got the download like last week. It was just unbelievable. What was great about it was that I could get a comprehensive understanding of my body, what's happening, but also somebody there reporting it out and talking to me through how we understand it, what to do next, and a plan, and basically build a plan from there.
It was great. I actually purchased it as well for my parents and things like this. I think it's just a great gift.
Yeah. Dave, we need to get you there too. Why Orlando and why not somewhere else?
I was on the East Coast, so I popped down to Orlando and so it was just easy for me.
So yeah, we got New York, Orlando, Naples, Dallas, Houston's opening, Miami, and LA. Anyway, back to the conversation here. I can imagine this is going to up the value of health in the home a lot, right? So one of my visions of the future is you're constantly being monitored for your blood biochemistries, what your protein levels, your vitamin levels and so forth, and that's being uploaded to Figure in the kitchen cooking your meals ideally suited for what you need in that moment and then the whole elder care side. How do you think about that?
Yeah, it's like, so I grew up on a farm in the Midwest and then my parents got into independent assisted living like 15 years ago. So I kind of grew up around senior care a little bit in my life.
You got into that business?
Yeah, my parents own and operate senior housing facilities in the Midwest.
So wait, they're still in Illinois?
Yeah. Wikipedia says your hometown has 2,000 people in it.
I grew up in Miko, Illinois, man. I think it was like 1,800 people when I grew up. And yeah, middle of nowhere. We had no traffic lights, no fast food. It's a dry town. It was just a whole different world.
Do they have parades for you when you go back home?
Man, it's just parades going through the street.
Can you imagine that? Yeah. So you understand the value of a fully autonomous human robot.
Yeah. We got to put, I'm really passionate about figuring out how to be able to ship robots into senior care and letting people age in place at home. Like even, you know, it's hard to get people to move into assisted independent living facilities.
Well, how's that work? So you're sold out, you know, 3 years into the future. You can't make them fast enough to keep up with the demand. And then you've got BMW, you've got a bunch of industrial use cases, but then you've got this in-home and you've got...
Really, I'll give you my, I'll level with you on how I think about things. We've been spending the last three and a half, we're about three and a half years old, trying to figure out what the right recipe is in the first instance of what a general-purpose architecture would look like for humanoids. We believe we found it internally and we understand what that is. And we believe we know how to make robots now and put them out and we're going to run them really hard this year. We're going to run them for...
You showed us, what do you call it? The grid.
Can you describe what we saw in...
The grid's like my favorite place here. It's one of, we have four buildings on campus. It's one of our buildings here and we have the facility outfitted that we're going to expand hundreds of robots into that will run 24/7. And it has a little mission command post, like a 007 situation room, and you can see every robot there and it's going to be doing both home and commercial workforce. We're spinning up, right now the facility just got open this week. You guys saw it's squeaky clean and we'll start shipping Figure 3s into it this month.
So model homes, model factories, model operations.
Well, so within mission control, you think of watching the robots, but the robots also have their own vision which transmits back. So it's more like in the combat movies where back at the home base, they're watching the invasion or whatever, you're seeing through the eyes of the soldiers. You got all that data coming back into mission control, too. So if the robot, there's 200, how many in there at any given time? A couple hundred.
