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.
And this was the first car in the world built by a humanoid robot that we're aware of. We have a robot here that is designed to tidy the house. These robots are running purely autonomously from an onboard AI policy called Helix. Give it a push.
Oh gosh. Okay. I feel bad.
Thanks. So where are we right now? What is this? This is the headquarters.
Robot campus. We make robots here.
We make humanoid robots here.
We design them and we build them and we test them all here.
So for people who don't know what a humanoid robot is, you will see in a second because this is kind of freaky, but can you explain what they are?
Yeah. Our goal is to build advanced AI that we can put into a general purpose humanoid body. A humanoid is basically just like a robot with a human form. So, we have arms, hands, head, 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 basically design and ship humanoid robots in the world that can do everything from housework to dishes, laundry, to manufacturing, healthcare, just basically as many things in the world as possible that we can go out and ship robots to. Oh, we see this on the screen.
Oh, yeah. That's our... This is your hype machine. Yeah.
Yeah, exactly. That's our latest generation robot, Figure 3, doing a little like concierge job.
Cool. 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.
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. These robots basically can run fully autonomously without any humans, and they can automatically dock themselves and charge. And then once they're fully charged, basically come off and be able to do useful work. So these robots that are docking here are charging through the feet. So we have a wireless charging stand, similar to how you would charge a phone inductively. The robots can charge through their feet at two kilowatts. So basically the battery lasts about four to five hours, and we can charge basically for an hour and go back and do work again. So the robots, we don't need to do anything. We don't need to plug them in. 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?
Yeah. They're just docking here just to charge and then they'll be out doing useful work all day.
Okay. And I saw these ones as well. 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.
We got a call asking to be basically the first humanoid group ever to put robots in the White House. And so last week we basically had the first humanoid robots in history there doing stuff, talking, greeting folks. We were basically at a very special event with the First Lady and it went really well.
Yeah. So this is our corporate headquarters. We have four buildings on campus. So here we do a lot of basically engineering design work.
How many people work out here?
We have about 500, a little over 500 people.
How many are in the total company?
Oh, sorry. There's about 250-300 people here and we have 500 in the company.
Yeah. I would say most all of it's engineering and then we've been growing out manufacturing, supply chain, some of the areas to basically how do we make more robots pretty aggressively.
And how many robots do you have here? How many are there?
Okay. Are they outpacing the humans or...
There's... my goal for this building is I want more humanoid robots than humans.
Okay. Walking around and they don't necessarily need to walk. They could be sitting or talking.
Yeah. So basically we do a lot of hardware and software validation testing here.
So like testing for burn-in, durability, basically like any new software hardware gets validated through this facility.
Yeah. So you have robots doing all kinds of crazy stuff here.
That is definitely some yoga. Robot getting on the ground and getting back up again. We're basically trying to stress test the robots to try to find any potential failures before any new hardware software could get released. So if we have a new camera, a new type of 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. The robot's basically made up of about 40 motors.
Every motor can spin like 360 degrees, all the way around.
So mathematically, how many states it could be in, like body positions?
Is 360 to the power of 40?
Yeah. It's more body positions than atoms in the universe, which is crazy.
Yeah, I've done the math. It's for sure.
You've done the math? Yes, for sure. Okay. So it's basically like the difficulty here is like how do you control it? You can't write code to make this work. So all of our robots here run on a neural network we call Helix. It's a vision language action model we designed here internally to tell the robot what to go do, to stay balanced, 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.
Yeah, they're phenomenal.
We basically have a large scale data collection effort that's going on and then we train our own models here internally and then we test them all here as well. It's not just for balancing and being able to have stability, which we need to have like human-like stability. It's for how do we know what to go do? 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 and logistics that you'll see here later today. And we have to do that like a few hundred times a second from camera images and on a neural network that runs on board the robot. So it's a really hard project.
Yeah. So yeah, all the bays here are running some sort of test for durability or reliability testing.
Why do some have different suits?
They have all the outfits are fabric, like human clothes. Yeah.
And they all have different clothes.
Which is cool because you can accessorize the robots how you want it. Our clients can have different outfits that show the client logos and the colors. Yeah.
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 without a technician.
Yeah. All the robots have a little zipper on the back. I'm sure you see. We can basically just unzip it and basically take it off and put something new on.
Yeah. Also, it's just really cool.
It is pretty cool. Their shoes look like real sneakers.
They're high tops. They're high-top sneakers. And yeah, they look awesome.
Yeah, I mean they look like human bodies. But getting the lab to this level of infrastructure to be able to run them like this every single day is actually quite difficult. And then we need to be very diligent about when we find issues, how to track them, how to do fault analysis really quickly and then how to solve them across the global fleet, wherever they're at.
How many are you 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 checks for all software that could be embedded software, could be a neural network like 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 we're trying to run a use case for logistics or home and the robots are messing up. We're not sure why it's messing up. That's not great for us. So basically here we're doing a ton of testing. We have test plans laid out every single morning. We're running those down and we see any potential faults, we have to go solve it. Then we have to retest those plans. So these robots running here all day every single week and we run them really hard.
And the goal is we don't want to be finding failures upstream out of this lab.
Yeah. So this is like the banner on here on system integration tests is trust but verify.
Okay. So yeah, we basically have to make sure we run down every potential thing that could go wrong before leaving here. This is a hard thing. The robot has 40 plus moving joints. It's like a walking cell phone self-driving car. All the supply chain is basically new. We've designed almost all of it. And so it's hard. It's hard to get the system to be really reliable. And it's not like if we lose power, we're not statically stable. So if we lose power the robot falls. So we can never lose power. We can never lose comms. 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 hand moving and things like this and head. So it's quite a difficult problem.
And so that's why you dock them at 15 minutes or 15%?
We dock like around 10 to 15%, they'll go to dock and then if we need another robot in we'll sub them in off the dock and you'll see here later in logistics and other use cases that need constant 24/7 attention. The robot will undock right before the other robot needs to go leave and the robot will basically do a quick swap and within like 30 seconds it's now doing work. Another robot will go into dock and we'll just run that every four or five hours on repeat 24/7.
What's the most common error that they make?
We mostly is software at this point.
Yeah. Okay. The hardware has gotten really robust. I mean, the hardware here is great. It's basically a software issue.
And then in terms of hardware, you're manufacturing also on campus.
Yeah, we manufacture here at BotQ next door.
Yeah. We're going to show you that today.
Yeah. Okay. So, we have a robot doing burpees here. We want to be able to safely for any event get down to the ground. And then we want to be able to safely get up. It's important in case we're not sure what to do. We could be on a very low battery and we might need to safely get down. And then we want to be able to get up off the ground really easily. It's also quite hard. You need a ton of range of motion in the legs and the hips to be able to do this kind of maneuver.
Yeah, he's going to 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.
And then we have some soft foam underneath the legs and arms right now.
Obviously the more padded and more soft I think is great. It just adds a different level. It adds more volume and mass to the robot.
Yeah. It makes the robot look bigger basically.
So, how often do you come into the office every day and you check in on them? Like, do they feel like your babies? Like, do you feel like they're for sure babies? Parasocial relationship with them.
They're like we made these like they're little kids and we have to get them to do useful things now. I think the good news is we're at a point where the hardware has gotten like very robust. We can run them all day every day. We still see hardware failures but 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. 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 without failures. Like most of our failures today are kind of in software land. And speaking of software, Mort's like one of our favorite things to do is push robots around.
And so Mortz here is one of our leads on the Helix controls team and maybe you can give her a quick 101 of the S0 controller we have here and then we would love for you to push the robot as hard as you can.
Yeah. So I think what we did recently switched fully to RL from a model stacked on this is that we have all this variation that we can give to the robot when we train it in sim. So all edge cases are now known to the controller and gets robust to it. So what this means for example before a model-based robot stack goes freaked out, you have this very robust to external services. We push it around. Go ahead to convince yourself really. I think this really nicely showcases how robust our stack is. Give it a push. Oh gosh. Okay. I feel bad. That's a little harder. Okay. So they don't... No, they like... what do they do if they get attacked? You know how like Waymos get attacked? People were attacking Bird scooters. Do they have a defense mechanism?
No, it's not trained into them. They don't see the internet. And no harm to humans.
They're here just to help.
Yeah. They're here to take a push too, if you need to.
Yeah. You want to get another one in?
Yeah. It feels really heavy.
Yeah. It's like 135 pounds, but it's actually got human-level stability. And as Mort mentioned, we're learning that coverage in a simulator. So the whole controller learns how to stay stable like this synthetically in a sim, basically like a video game. And from there the robot learns how to stay balanced, how to not fall whenever there's certain forces and we basically can zero-shot it onto this robot meaning we can just load it to the computer and get this level of performance in the controller.
What are like again, what are the most common tweaks within that in the software? Like you're obviously balancing a lot of physical issues there. So how do you tweak that in the models? Mortz, how do we get the models to be able to perform like this?
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. I think that's the... we basically have like a physics simulator. So it has gravity, it has friction coefficients. We want to try to mirror what forces robots are seeing here so we can run them in sim. 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 in a simulator like a video game, it can stay stable with these forces. Then we load it to the robot and we see the same in the real world and we have basically a very high transfer rate.
It's interesting. I feel like it would be interesting to hear your perspective or differentiation on your robot versus the other humanoid robots and where this physicality and essentially the behavioral mechanisms change. Some might be more commercially focused. These ones are clearly as humanoid as I've seen.
Yeah, we're a pure-play humanoid. Okay. Yeah. No. I think a few things. 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 my iPhone has a bunch of different apps and if I wanted to learn something new, I just download another app. The same thing for humanoids. We want 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. And so we want one set of hardware that we can advertise over a lot of different use cases. So the goal is to get the hardware robust enough to do most things a human can in terms of range of motion, be able to get down, get off the ground, be able to reach up high, be able to reach inside of a sink, all these different things we need to do. We also need to carry a decent amount of payloads and we operate at decently fast speeds. So we've designed the hardware to be able to do that. 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 and then reason through pixel space, take camera videos of what's happening and then output what the body should be doing. The math we were doing before on how many states the robot could be in is just so high.
That the problem just kind of runs away. It's like a curse of dimensionality. It's just too hard. So you can't solve it with writing lines of code traditionally like in robotics. And even three or four years ago here at Figure we would solve the same controller here in code. We'd have hundreds of thousands of lines of C++ to figure out how to solve this inverted pendulum math of how to stay balanced here. And what we found is it just doesn't scale, doesn't work. You can't really in your head as humans work across many different humans to code all the stuff into the robot that you think it could encounter in the world. It's just too difficult. So we transitioned purely to neural networks with Helix, our neural network model.
For those... I know you mentioned this a couple times but this will be like 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. We trained a transformer policy to basically output a certain type of action space. Basically we train an AI policy to do this work of what code used to do and learn this. So now basically we run inference on board the robot 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 similar to how you'll talk to an LLM and you'll ask it what to go do. It'll do inference and output basically a next token prediction for words. We do the same thing here but for a physical humanoid. And we'll output things like where to put the wrist, head, torso, every joint will get an output from the neural network like what to go do. And we'll do that anywhere between 50 and 200 milliseconds. So basically like 50 to 200 times a second, the neural network computer is then telling all the joints what to go do. And then every joint level, our motor controllers are outputting torque on where to position the motors.
So yeah. So basically you can have two paths here. You can basically code your way out of this and I think that's a full dead end or you can run an AI-first strategy in a market and that's what we do here at Figure. So I think what differentiates us is from a hardware perspective I think we probably have the best humanoid hardware in the world to do general purpose work 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. So you'll see some use cases today that we do both for the home and for commercial market. Those are all run by our Helix neural network which is hard. We have to tell the whole body how to reason over camera frames and then how to position itself fast and dynamic. So you'll see in cases like logistics these packages are moving while we're grabbing them because they're plastic. We have to reason over all of that in real time and be able to do like if things move in real time we have to basically closed loop react to it. That's really hard. We've been spending the greater of the last two years trying to solve this problem. And I think we probably have some of the best AI policies for humanoids in 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 US. Nearly 40% of startups fail 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. Their all-in-one solution combines checking, 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, ServiceTitan, Hims, Anthropic, Flexport, Robinhood, and Plaid trust and use Brex. Start today at brex.com/sorcery. That's brex.com/sorcery. Turing is training the next generation of AI with tasks that require real expertise and real-world judgment. That's why companies like Nvidia, Anthropic, Salesforce, and Gemini partner with Turing. Turing builds realistic reinforcement learning environments and data systems based on real operational traces. The kind of infrastructure frontier labs need to train superintelligence. Visit turing.com/sorcery.
Another thing I want to show you is this is like we just walked out of system integration and test the lab. We're trying to find all these different failure cases.
One of our failure cases is for a humanoid like we're balancing. So if we lose power the robot falls. It's the same for losing power or losing a motor in the leg.
Do they need to be connected to Wi-Fi too? What are all the integrations?
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 internet connection or network, they can basically continue to do work. We run Helix on board. So they're actually loaded into GPUs on board in memory and we run inference on board. Meaning if we lose internet connection, we can still do housework and logistics like humans are.
I mean maybe I have a hard time doing work when I lose connection, but most humans can do most work. So another thing that we're working on solving that I'm excited to show you here is what if you lose communications with any of the 40 joints or what if you lose power? And upper body is kind of fine if you lose a wrist or elbow like you're not going to fall at the very least. But falling is a terrible event for us. We don't ever want to fall. We actually have an initiative internally called Never Fall. It's like we never ever want to fall even. And you know we will fall but we don't ever want to tolerate it. The hardest problem here for the controller is like what if we lose an ankle, a knee or hip? What if you just lost a knee? Normally for humanoids you just literally fall over like you can't really balance if you lose a knee. We've been working on a project we call Vulcan here internally that basically allows us to lose a single or even multiple joints in the legs and still not fall. So what we're going to do here is Mort, can you show her what would happen if we lost a left knee? So here we have a view of all the different joints on the robot. Green means we have comms and power to him and the robot can basically communicate and it's fine. And red here will mean we'll lose certain communications with the robot. So here we lost a left knee and you can see the robot's kind of like hobbling on the left leg. Yeah.
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 locked the knee and we can basically hobble around.
I thought it was going to be a little bit more dramatic. Honestly, it's not bad, right? It's not that bad.
What's unbelievable is that we can even do this kind of work. We're doing this also inside of a reinforcement learned neural net controller. So the same stuff that Mort talked about earlier of us learning in simulation, the robot learned in simulation how to move the body here to extend it lost different joints. This is... I watched the robot a few weeks ago. We were doing work on this basically logistics use case and it's like months ago we'd lose a knee. The robot would just fall. Now the robot loses a knee. It can either continue to do work or can just hobble off and another... I 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 and we basically don't lose any time.
Yeah, it's really impressive. I think this is kudos to the controls team. I think we have one of the best controls teams in the world. About a year ago, we made a strong pivot away from code into neural networks. And I think the team has probably shown some of the best neural networks for control in the world on humanoids.
What happens if it loses its knee while it's bent?
It'll basically do a velocity lock on that knee and should be able to hobble around it. Obviously it depends where. If we're in a full squat down, yeah, that might be really tough. It depends what state it's in, but in most cases we can basically recover from this at this point and survive. You want to build a real-time operating system that is basically fault-free. So this team here is all focused on trying to figure out all of the potential faults and risks. This team here is responsible for building our controller which is responsible for how do we move all the joints on the robot, keep balance and ultimately end up doing tasks.
How many different teams are in the company and how do you portion out who works on what and who gets integrated?
So this is part of our AI team. It's one of our biggest teams here. We have multiple different groups inside the AI division. We also have a team that does hardware. So they design motors, batteries, wire harness, structure, joints, kinematics, basically a wide range of stuff. We have a platform software team that does all of our embedded software, firmware. We also have an electronics team that basically builds all the PCBs and electronics work that we do here internally. We design almost... there's over 100 PCBs that we design here on the robot that we do on that team. Things like our motor controllers, all this different type of work. We have a system integration test team that you saw today that helps basically make sure we ship really good robots out to the world. 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... how do we design something that people really want? And I think we have a pretty high design standard here of designing something that is a really delightful piece of technology. We also have teams that work on system design and thermals. 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 BotQ team for manufacturing, supply chain team, facilities, and then we have all of our business operations side. So maybe like a dozen teams here internally that are needed to do this. It's basically the same teams that you need to build robots. So like any kind of robots, like when I was at Archer building aircraft, it's like a flying robot. So you have the same type of stuff. You have electric motors, batteries, control software, embedded systems and sensors. So we have basically teams around all that. And then obviously on the AI side is a big focus here internally. Do you want to see some stuff in the home?
All right, let's go. Thanks, Mortz.
Thank you.
All right, so I was hoping it would be more dramatic, but lose the knee.
Yeah. It's honestly 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 a Never Fall team and it's a team that basically predicts any potential faults and designs around them. This is a case where I think you're always going to have a period of time where you're going to lose a lower body motor or actuator. Then how do we survive this?
Okay. Outside of stuff we're doing on the commercial market, one of our big focuses here is how do we ship a general purpose robot to do things like in the home. So, all right, come over here. We have a robot here that is designed to tidy the house. So clean up any... clean up the table, basically put away all the different cups, clean the toys up, tidy the couch. Things in my house is kind of chaos. It's spraying and cleaning the table.
Don't you want... We all need this. It's like...
It was. So what's cool here is the robot is running an onboard Helix 2 neural network to be able to do all this work, to be able to do all this cleaning. So it's basically just taking a prompt which is like clean the living room and it's basically reasoning through what to do from its cameras and it's ultimately telling the whole body what to do from a neural network.
How many hours has this been trained on? Doesn't it take millions of hours to train?
We've probably had in Helix like the base pre and mid-training Helix models that we're working on now have maybe a little under a million hours of total data. And then we also do from mid-training and pre-training we do post-training 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 neural network platform that can basically do things like tidy this living room but also do things like logistics.
And this isn't teleoperated.
This is not teleoperated.
There are rumors that these are teleoperated.
For sure not teleoperated.
You don't have a secret room back here. Who's in there?
No secret room. These robots are running purely autonomously from an onboard AI policy called Helix running on board the robot in the torso. And so this robot's 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.
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 this same humanoid can go over and do logistics and healthcare and manufacturing or do dishes like 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.
When they're in people's homes, will that 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 large scale deployment is data. We need an enormous amount of data. We need to pull a lot of our resources further into pre-training for the Helix team 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 and we do want to be sending that up training on it into 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 fully want to anonymize all of it. 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 seeing 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 US now because if you go to Europe there's obviously...
Most of all of our work today is in the US.
