A Humanoid Robot in Every Home? It's Closer Than You Think w/ Brett Adcock (at A360 2025) | EP #156

Thank you for being here. I know that with three young kids, a robot factory in production, and an incredible team of engineers, you're really busy, and I don't take it for granted that you joined us here.

Yeah, thanks for having me. My only request is that next time I want a Figure robot with you.

Loud and clear. I begged him, and BMW has been taking the lion's share of them.

Yep, we do have a lot. We actually have them running every day now. They're there today, running in their largest plant.

Why did you start Figure? You had a few incredible successes. Archer was amazing, and then you jumped into arguably what could be described as one of the most difficult businesses to get into.

I think we really need to figure out a way to give AGI a body. I think it's a really negative, almost dystopian future if we figure out how to solve AGI and it lives in a server somewhere, and it's more intelligent than all of humanity. Ultimately, if it wants to do something in the physical world, it will have to ask a human to do it.

The humanoid robot is the ultimate deployment vector for AGI. You can't solve this with anything else besides a mechanical human. You need something that is a single platform that, with no hardware changes, can do everything a human can, and you need something that can also be good for neural nets.

A neural net in a humanoid can basically learn from transfer learning. It can multitask across a variety of different applications, which is really good for a neural net. We can build one single neural-net foundation model that can empower the whole robot to do everything end to end.

Massive congratulations. You went from a cold start in 31 months to shipping your first robot, which is extraordinary. A lot of companies get their PowerPoint decks ready and raise their first capital in that period of time.

We're going to be seeing some of the robots in the back here. When I visited you up north, you showed me around; we did a podcast together, and you showed me Figure 1. Here's Figure 2, and here are the designs for Figure 3. One of the things I truly find amazing is the speed of your iteration. Can you speak to that and how important rapid iteration in hardware is? Because hardware is hard.

This is a hard problem. We have to figure out how to do something that's never been done before, and it's a very complex system—definitely more complex from an engineering perspective than Archer was, building an electric aircraft.

My rule of thumb is that the first or second generation of hardware is always going to suck. The first iPhone was not great. The first time you make something, you're never going to get it right in hardware. You have to see 5 years into the future, know exactly what the product does, and then clean-sheet design it for that exact thing on day 1.

If you mess up any of those things, you can't go back and fix them through the design process. You have long-lead-time supply chains and everything else, so we are designing a new hardware platform every 12 to 18 months.

By the way, that's pretty amazing just to hear: every 12 to 18 months, a brand-new iteration.

Yeah, we had Figure 1 walking. By the time I filed the C Corp, we had the robot walking in under 12 months.

Another thing you've done is completely vertically integrated. Was that a necessity? There was no supply chain for humanoid robots. There were no motor vendors, actuator vendors, sensors, battery systems, structures, kinematics, or any of the software, which is pretty vast: firmware, embedded systems, operating systems, middleware, controls, and AI.

Walk me through your factory. You walked me through it before, but what are the different segments? What's going on there?

In terms of design, we clean-sheet design everything from basically the ground up. All the hardware is clean-sheet design. We look at what the product ultimately needs to do. You want to talk to a robot and have it just do things without any human intervention. You want it to go out and do things in the world.

We're designing for a capable robot that can do everything from working in a home—walking the dog, making coffee, and doing the laundry—to working in the commercial workforce, which is roughly half of GDP and is human labor. That's the largest market in the world.

We do all the hardware design, including kinematic design, joints, motors, battery systems, and sensors. We do all the software, firmware, embedded systems, controls, and all the AI work end to end. Then we do all the testing, manufacturing, integration, fleet operations, and deliver the robots to the clients.

We have robots now and 2 commercial customers. The first was BMW. We have robots there operating every single day in Spartanburg, South Carolina, helping to build cars. We have a second customer we just signed, and within 30 days of starting the work, we were doing everything end to end with neural nets. This is one of the largest logistics companies in the world. We're also pushing really hard on the home.

The global GDP is $110 to $120 trillion. Your TAM is like $50 to $60 trillion. That's pretty good.

Yeah, it's going to build the biggest business in the world by a long shot in our lifetime.

We have some video from the BMW plant. If we can roll it in the background or repeat that video, we'll show it.

This is a quick update from BMW. We have robots that are basically putting sheet metal on fixtures. This is a job that every major manufacturing company in the world does. Our robots are doing that fully autonomously, at the speeds we need to hit high performance, with no human intervention, no faults, no failures, and no drug testing.

No days off.

No days off. Twenty-four hours a day, 7 days a week.

I mean, it's an interesting thing. Let me jump into one thing on volume in the future. I believe I heard you say that you'll see these at a price point of $20,000 to $30,000. Do you still hold that?

We've done a lot of work on the bill of materials. If you break this down and look at it line item by line item, and what it looks like in high-rate manufacturing, there's really nothing in the system right now that would show that this product should be extremely expensive.

The calculation I do is that if I were going to lease a $30,000 car, it would be about $300 a month, which is, by the way, $10 a day and $0.40 an hour. So here's my question: How many of these humanoid robots would you own at $300 a month, operating 24/7, with no complaints and no fights with girlfriends or boyfriends? The number could well be multiple robots per human.

You're going to want one. I wake up every morning and help unload the dishwasher and pick up the kids' toys. I never want to do any of that ever again. It's not something I need to be doing when I get home or when I'm at the house.

We really haven't had a lot of innovation in the home for 50 to 70 years. We have the same appliances and the same stuff. We had old robots—we called them dishwashers—and they've been around for a long time. As humans, we're having to work with those machines every day, and that's not something you'll have to do anymore in the future.

You'll just talk to the robot and have it do it. It will be on a schedule, and at any moment you can call it, text it, or talk to it and ask it to do something. It will know you better than you know yourself.

I remember a couple of years ago—I’m very proud that Bold was an early investor in Figure—and I brought Tori to meet you. I said, “Listen, first of all, Brett's an incredible operator with multiple successes. What's one of the best predictors of the future? It's what a person has done in the past.” That's very much one of the best predictors.

But what I found amazing, beyond your charm, that sold me instantly was the team you pulled together. Can you talk about that? A lot of people in the audience are focused on their moonshots, and this very much is a moonshot. You exited Archer—how did you capitalize, what did you start with, and how did you pull your team together? Describe that early moment.

I haven't founded a lot of companies in my lifetime, so I get to go back every time and ask, “What did I mess up on? What did I get right?” Then I try to make things better.

Fundamentally, in order to build one of the world's greatest products, you need one of the world's greatest teams. You need to align that team with the shared vision. Everybody needs to be accountable for it and understand it, and then you have to figure out how to hit the gas pedal really hard.

The entire culture at Figure—even at Archer, when I built the initial team—was very deliberate. If you go to the Figure website now, we have the culture deck, the master plan, and things laid out that are really unique. We're in Silicon Valley, but we're almost the anti-Silicon Valley: You have to work every day in the office. We work 5 to 7 days a week, and we work really hard.

Not a lot of people want to do that, and that's fine. They're just not the right people for us. We've assembled a couple hundred of the best engineers in AI and robotics in the world. There's just nobody even close to what we've done.

My whole business team has been with me at Vettery, Archer, and now Figure. We've spent 15 years together. They're unbelievable operators. They give me the ability to spend basically all my time on product engineering to build the best product possible, and they help scale the business, which is great.

Hiring, recruiting, HR, legal, finance—across the board, they're great. The team is insane, but what's even better is that the culture is absolutely dialed in. Everybody knows what they should be doing. I don't do 1-on-1s or things like that. We have a shared vision of what to do, and we work really hard to get there.

The dopamine that we all get is the same. We want to ship product, and that's what we're aligned to. That's what everybody gets their dopamine from, which is really great. It's this shared fuel that we have to ship product.

This humanoid work is one of the most complex things I could have worked on. You fundamentally have to have that, or there's literally zero chance this is going to work.

We're going to hear from Travis Kalanick tomorrow, and he's going to say very much the same thing: Your massive transformative purpose, that clear mission and vision, and then aligning your team and culture around it. It starts with you.

You made a commitment of your own capital to get it going, and then you started calling people at other companies. What was your pitch to raise capital?

What's that—to raise capital or recruit?

No, no, to get those employees on board.

The pitch in 2022 was, “I'm going to fund this whole thing for many years.” We got to $1 million a month of burn in 6 months, so it wasn't cheap. But I was full pedal to the metal from day 1. I knew exactly what to do.

Archer is kind of like a flying robot in a lot of ways, so I knew how to build teams. I knew how to move quickly, and I had the technical understanding of powertrains, control systems, beta software, and sensors. We moved really quickly from there.

The pitch was, “I'm going to fund it, so there's no funding risk, at least in the near term—the next couple of years. There's a good chance for us to build the next iPhone moment happening with humanoids. It's going to happen right now.”

What did you tell them the probability of success was?

Pretty low. We needed to prove 3 things that had never been done before, and we had to get all 3 of them right in under 5 years, or we would fail for sure.

First, you have to build incredible hardware for humanoids that's extremely complex. It can never fail. It's always got to work, and it's got to work at human speeds with a human range of motion. Nobody's ever done that before. Most robots that walk around can't even walk properly; they fall over all the time. It's maybe rocket- or turbofan-level complexity in terms of hardware systems.

Second, this is a neural-net problem, not a control problem. You can't code your way out of this. You can't hire a PhD with a robot and solve every problem. You have to ingest human-like data into the robot through a neural net, and it has to be able to imitate what humans do. This has never been solved on a humanoid system. It's a high-dimensionality system, not like a robot arm on a table, most of which don't have AI.

The third thing you have to do is figure out how to generalize. That's the holy grail of robotics. You have to figure out how to look at something you've never seen before, tell the robot through speech how to do it, and then have it execute that task fully end to end with one neural net.

I wrote about this in the master plan in 2022. We needed to solve those problems. If you can solve them, you're in the right decade. You're going to build the iPhone moment for this whole space, and we're in full liftoff.

Those looked pretty dire at the time. In 2022, there was just nothing out there. You had Boston Dynamics leaping around, doing backflips and parkour, but nowhere near the level of manipulation and dexterity you needed for humanoid robots to enter the home.

I think we can confidently say now that we've solved—or are making substantial progress on—all of those.

Amazing.

There was a pivotal moment late last year when you said OpenAI was a large investor and you were baselining OpenAI's AI systems. You made the critical decision that we had to build our own AI internally: Helix. Can you speak to that moment? I'd like to show the video of Figure at home along those lines.

What you're seeing is Helix. This is our large-scale AI internally. It's basically a large-scale vision-language-action model, and this is public; it's on our YouTube channel.

The prompt that Cory, who leads the Helix team, gave was, “Put the groceries on the table.” The prompt was simply, “Put the groceries away.” It didn't tell the robot where they go or what they are—just to put them away.

The tricky part for the robots is that they had never seen any of the groceries before in training. We purposely withheld all of these items, so this was the first time the robots had ever seen them in their lives through their own cameras and sensors.

You have to solve the generalization problem in a home. Every home is different. We all have different toaster ovens, appliances, spatulas, and silverware, and everything is located differently. Things are also changing throughout the day.

You really have to solve what I call semantic intelligence, or semantic grounding, from the human world to the robot world. Helix, which we can talk about, is able to communicate on a single neural net on each robot and collectively put all of these things away with a single English prompt.

I think this is the first sign of life. I’ll go even further with a bolder claim: I think this is probably the most important AI update for robotics in human history. Everything in the future that moves will be a robot, and it will be powered by AI agents like this.

This was also trained on very little data—500 hours of data. Peter Diamandis

I love the way they're looking at each other to confirm, “Yes, I get it. Where are you putting that thing?” Yeah, I think that's a good idea to put it up there. They're about to look at each other here as one robot passes the item over, almost as if to say, “I get it.” Is that created?

Part of this was emerging from training. When the robots are doing handoffs, there’s actually a split second when one robot needs to release the item and the other robot needs to grab it, so it doesn't lose hold of the item and drop it.

What emerged from training was that the robots actually look at each other as a clearance signal that they should release the item into each other's hands, which was really interesting. The other aspect—robots looking at each other and moving around—is important overall.

There's a certain level of communication that needs to happen from a robot in terms of interaction design with humans. You don't want to walk into a room and have a robot just not move or not look at you. Humans look at each other and use nods and gestures. All of this is extremely important to learn.

We need to learn human expressions just as we need to learn how to grab items. It's going to be super important as we integrate robots into the entire world at scale that this happens.

I have 1,000 questions for you. Let me hit a few rapid-fire style here. Figure 3: When do I get to see it? I saw the designs.

Yeah, you keep asking about this one. You like this one. You saw it.

I mean, the degree of beauty was increasing.

I don't think people understand how incredible it is. They don't, because we haven't shown it. Figure 1 was online a little bit, but it was more gnarly, with wires outside of it. It was faster, and it was a much quicker design cycle to get something to our engineers so they could start doing real use-case work.

Figure 2 was a feature-complete robot that was supposed to be able to do almost anything a human can do, or the vast majority of it. We haven't talked about this publicly a lot, but we're done with the Figure 3 design. We'll probably show an update next week—just a minor update, not anything material—as it relates to how we're going about that process.

If you look at Figure 1 to Figure 2, it's a huge step up. You go from a college dorm-room project to a real, pretty decent robot, and the magnitude of that step was substantial. That same magnitude of step happened again with Figure 3.

If you were to see it, it's unbelievable. We spent 18 months designing it from scratch. At a high level, it's 90% cheaper, it's smaller, it has less mass, it has better sensors, and its hands, head, and feet were designed for neural nets.

It's a completely different level of design. Figure 2 is probably the best humanoid on the market—maybe not by a lot, but I think it's the best by 10% to 20%. Figure 3 is next-level design. For me, it's the proudest moment I've had in engineering in my career, looking at that robot.

We're going into production manufacturing with it this year. We'll have more updates on that soon. That's the robot we want to send everywhere into the world. We want to make it low-cost and very high-rate, and it's better across so many dimensions.

Tell me about production rates over the next 3 to 4 years, and when am I going to see it in the home?

We have 2 tracks: the workforce track and the home track. What most people don't understand is that the workforce is the big business. It's half of GDP, we can charge meaningfully more per robot than we can in the home, and it's also easier. The things the robot does are almost the same things on repeat.

The home is the Wild West. It's extremely hard. We have a huge safety requirement around not falling on or hurting any human. There's also semantic safety, such as not knocking over a candle and burning the house down.

The home is vastly harder. In self-driving, driving on the highway is like the workforce for us, and driving into the city is like the home. It's unbelievably difficult.

Between our 2 first commercial customers, which are very large businesses, we have demand. If we had 100,000 robots today that all worked, they would take 100,000 robots today. I could sign 50 customers by the weekend, all Fortune 100 companies that we've literally visited and know. We just can't supply them.

I've had a bunch of meetings today over lunch. Everybody is asking what we think about helping out in healthcare and construction. It all sounds great. We're just bombarded by the amount of demand here.

When you think about the workforce, you have a certain supply of humans, and it's literally going down demographically. Baby boomers are retiring, so you have fewer humans in the workforce. There are labor pains everywhere, and there are a lot of job shortages.

We see unbounded demand. I think we could ship 1 million robots this month if we had them all working and ready to go.

One thing we're going to add before you go—sorry, I know you want to rapid-fire—but you saw BMW and you saw our second commercial customer. It took us a year to do BMW fully end to end at high speeds. Last summer, if you look at Figure 01, it took 4 minutes. Now we’ve gotten down to 40 seconds through a lot of great engineering work.

We started working on Helix, and it was completely transformative. Then we said, “What if we use Helix for this next use case for the new second customer?” We did that whole thing end to end in under 30 days, starting from nothing. If we had to do it all over again, we could maybe do it in less than 48 hours.

The robots are going to learn how to do something in a matter of hours—not 10 years from now, but this year. I think that has pushed our timeline for the home forward by multiple years. The long pole in the tent for the home is semantic intelligence: understanding what the hell is going on wherever the robot goes.

We’ll start alpha testing in the home this year. That means we'll be doing internal work in homes—my home or our engineers' homes.

You want to get rid of that dishwashing duty, dude?

I can't do it anymore. It's just not something I want to do. I want to spend time with my family, kids, and wife. It's just no bueno. So, yeah, we have to fix that.

At this point, we feel data-bound in the home. We think that if we increased the data set we trained Helix with by a couple of orders of magnitude, it would probably work.

Right now, Helix can pick up almost every small household object we put in front of it. We put a weird cactus toy from one of the kids' rooms in front of it and said, “Pick up the desert item.” It had to relate a cactus to a desert plant, even though it was a toy that was singing and moving, and it picked it up.

All of that is in the weights. It has a very large language-model backbone, so it really understands the world through semantic grounding. We think we just need more data now. We're basically data-bound.

I guess there's a lot of confidence that you're seeing a sign of life now that you haven't seen before: an intelligent robot in the world can be built. The question is whether we just have to keep extrapolating that curve far enough for it to enter the home.

I think it's this decade. You're going to see it in homes in the coming years. Just through speech, you'll be able to have it do very long-horizon hours of work without any prompt or any fix.