The Robot Episode: Four Leaders on What's Coming

All right, everybody. Our interviews with the number-one companies in robotics today continue here in Paris. I'm really excited to have Dr. Péter Fankhauser on the program. You're the co-founder and CEO of ANYbotics. You make the ANYmal—get it? A lot of you have puns.

You've been working in this space for close to 20 years. The company's been around for 10. The first 5 years were kind of a research lab. The last 5 years, what do you call these dog-based robots?

Well, it's an inspection solution, right? It's about data collection and understanding in critical infrastructure.

But the form factor is—

A four-legged robot, or a dog, as we like to call it—a dog robot.

Why did that dog format become the standard? You're not the only person making it. There are many people making it now. Why did that one become the first one to reach relative scale and deployment?

In nature, a lot of animals have 4 legs, so there's a reason for that. For sure, you have great mobility. You can climb stairs. You can go anywhere a person can go. So, dexterity and balance—

Mobility, right?

Balance, but also stability. With 4 legs, if you have a wide footprint, you have a lot of footholds to hold on to, because we work in nasty environments: slippery floors, rain falling, snow falling down, grass growing. So, 4 legs is a really good format.

Well, now this is a silly question, but why don't we make centaurs for the humanoid versions? When people are making the Optimus, the NEO, and the Atlas from Boston Dynamics, those stand-up robots have 2 legs. The concern is that they're always going to fall over. They constantly fall over in demos, and if they fall over, they're going to break somebody's ankle. Why not put 4 legs on those?

You absolutely could, and it really depends on the use case. If you need it to work in a coffee shop, there are narrower spaces, right? You want to work at eye level. Maybe a humanoid is better. In the facilities where we work, 4-leg stability is ideal, and there's enough space to go around. It's the perfect format.

I don't buy it. I think all the cafés should have centaurs. Were they centaurs in Greek mythology?

It's a centaur.

I think that should be the new standard. You found a really effective first use case, which is inspecting really important infrastructure, and now you have hundreds of these deployed over the last 5 years.

Yeah.

These are expensive. They're in the low hundreds of thousands of dollars to buy.

Yeah.

And to operate them, I'm assuming tens of thousands a year in service contracts.

So, they're not for home use.

These are industrial, and they have a lot of sensors on them. If you were going to inspect, I don't know, a pipeline with natural gas in it—

Right.

These things can go out in any weather, and they can sense things on that pipeline that a human can't. Correct?

That's right. For us, it's not about labor replacement, right? It's about what we can do better and what we can do superhuman. Inspection is a great example. Our eyes and ears don't perceive all the signals: micro-gas leakages, temperature, equipment overheating. The cameras on the robot include thermal cameras, acoustic microphones, and gas-concentration sensors. We pack it full of sensors and AI, and you can go way beyond what a human can do.

The monetary benefit is avoiding downtime. If these assets stop, they lose hundreds of thousands in revenue per hour. So, every minute or every hour we can save them essentially pays for the robots. That's why we can afford to have really expensive sensors and really expensive GPUs on top of a robot.

These have seriously powerful compute on them, right? And they have to have a significant amount of battery power, then. So, these things can do a mission for what, an hour or 2?

2 hours, and a docking station to come back and charge. But they do this over and over. Some of our customers run these missions 40 times a day, 24/7, because they're interested in a specific point when the electric arc furnace goes up. They want to know what's happening at that minute.

It's too dangerous to send in a person. The thermal cameras would get burned. They need a robot right at that moment.

Exactly.

And they have to charge—not hot-swap the batteries.

No, you want hands-free autonomy. Nobody should even be bothered that there's a robot. They don't care about the robot. Actually, they don't even want the robot. They want the data. They want the insights. The robot is a means to an end to collect the data precisely.

At what point can you offload the very power-hungry compute and put it in the cloud?

We also do that. There's always 2 parts. There are parts that need to run in real time on the robot, because you also cannot guarantee connectivity. Obstacle avoidance and data quality—making sure you have the right thing—need to happen on the robot. If you upload a blurry image to the cloud, it's too late.

But in the cloud, of course, you do contextual analysis, historical downtime analysis, et cetera. Are people asking for these to be able to operate for 24 hours yet, or 12 hours?

No, for sure. The maximum is in the 8-hour range, so it has enough time for charging. If you need to go beyond that, that's rare. There's diminishing returns to doing it more frequently.

But you have to manage it. They do it manually today, maybe once or twice a day, and they do it 8, 15, or 20 times now, right? So, the frequency already goes massively up.

Yeah. Without putting people into harm's way, plus the quality is so much higher. What's the most fascinating science-fiction deployment you currently have with these?

What's really exciting is anything offshore, right? People fly out with helicopters. Every helicopter flight costs in the tens of thousands. If you're offshore, it's very tricky. It needs to work, and there are almost no people around.

These are oil rigs?

Oil and wind energy offshore as well.

Ah, yes. But wait a second. These things don't operate in the water, so how do they work with windmills in the ocean?

There are hundreds of windmills around, and they come together at a transformer station. It transforms AC to DC before it transmits, and that's a manned facility, typically a big converter. This is where the robot operates.

Got it. Can they operate in severe conditions, like the Antarctic and stuff like that? Have you deployed them there yet?

In Norway, for sure. That's −20°C in Norway, and in deserts, plus 40, 50, or 60°C. That's exactly the point where you want to send in a robot: extreme temperatures, dust, and humidity.

Most importantly, we now have a robot that goes into explosive atmospheres where, in oil and gas and chemicals, there's methane in the air. You're not allowed to create a spark. So, we built a special robot that's guaranteed not to create a spark.

This is where you don't want to have people, but for a machine, that's a perfect case, right? A dangerous environment. This is where we're sending robots in.

So, if you're in the Permian Basin and something's leaking, that's one of the most dangerous situations—these oil rigs and gas leaks. This is where people seriously die.

Yes. You want to know when it's happening, but you don't want to create a problem. So, that's a perfect case.

I'm going to keep going sci-fi, but dropping these things to the bottom of the ocean seems like a no-brainer at some point.

Well, there are submarines, right? We don't do that right now, but I agree. Robots should work in environments where it would be dangerous for people—remotely, on boring, repetitive tasks. This is what we—

That's a different form factor right now. But there are people creating robots on the surface and then under the surface—slightly under the surface—that are doing, essentially, not inspections but monitoring systems, obviously for the military.

If you're out there inspecting and there's a gas leak, and it's dangerous to send humans out there, when are you going to put some equipment on these to fix the goddamn leak while you're out there? That must be the holy grail, is it not?

Yeah. Once you can detect a problem, customers ask, “Can you solve it? Can you fix it? Can you turn it?” Not today. In a demo, yes, but in reality, getting to 99.9% reliability in an explosive atmosphere is still in development.

The first step is closing levers and opening cabinets. Eventually, you want to have, through manual manipulation, maybe 3 or 4 arms to fix the machine. That's still a lot of work. AI will help us there, but there's still a lot of work ahead of us.

A lot of the demos you see of humanoids folding laundry are in a very controlled environment. Once you're outdoors in a hailstorm or freezing temperatures, it's different, also for perception. But eventually, we foresee a future in which this will be solved.

What percentage of your robot is sourced from China?

Zero.

0%.

Yeah.

Is that because in the EU and Norway it's banned, or is that a choice?

That happened just historically. We source locally, and you get chips from the US, et cetera. For some of our customers, it's important, and we built a lot ourselves because we started 10 years ago.

Nowadays, you can get cheaper components from around the globe. So it's about being smart about where you get components from, which ones are active and which ones are just metal. It's a hard world to navigate, but tapping into the commoditization of certain hardware makes sense for us cost-wise.

Who's specializing in that outside of China now? Is it Vietnam, India, Taiwan? Where can you source the actuators and a lot of this?

For sure, China is number 1 in pushing this. There are good companies in Europe and in the US as well. These 3 regions, for sure. If it's just about labor and assembly, you can go elsewhere as well, but you want to get the core expertise from somebody who builds that component.

Got it. How do you look at China now? They've been stealing intellectual property. I'm assuming they've stolen yours already. Certainly, other people's intellectual property is being stolen at scale in China, and they're building robots that are going to be 80% cheaper. They're going to try to deploy them to the same customer base, I am certain. How are you thinking about the threat of Chinese robotics?

If you look at the robot from China today, that device is a piece of hardware that can walk beautifully. Great engineering. I love it. It can do backflips.

But they're not solving the problem. Our customers don't compare a platform to the full solution that we have. You need autonomy, inspection, intelligence, workflow integration, and so much more. It's just a hardware difference.

So the harness, the wrapper, and the services around it—they're not providing that?

And then there's the trust in the data, right? We handle very sensitive data. We have ISO certification for cybersecurity and all these topics. That's how we compete.

So you might not want to send the latest data from a nuclear power plant to the Chinese Communist Party, you're saying?

You don't want to have 15 cameras in your critical infrastructure that somebody else controls.

Yeah, I'm being a bit facetious, but—

It's happening today.

Yeah, but there's data leakage. Talk to me about military applications.

Yeah.

NATO is having to arm itself. I apologize on behalf of the United States for our stance with NATO, but you guys have to pay up and pay your fair share. You've agreed to do that, but I think there's a perception in Europe—you can tell me if I'm wrong—and in NATO that you may have to go it alone, maybe without the United States. You may need to build your own military products and services.

Do you need to be in the military space? Do you need to take the same applications and build military applications? Are you doing that yet?

Yeah. So I think there's a responsibility in Europe to build technologies to be able to—

You believe that personally?

Yes. However, for ANYbotics, we built and went down one track. There's tremendous pull. So today, we're not doing it, and we don't intend to do it. It's also a different product at that stage, probably.

It sounds very easy: just take 4 legs and do military. You need to go a couple of steps further. What exactly are you doing differently? Different communications, different autonomy. So we're not doing it, but I think there's a responsibility to do it for others.

Is it “never say never” for you, or are you dead set on the idea that you have a mission and you're not going to build a military product?

For us today, the mission is clear. We started with nonmilitary applications, and this is where we're headed.

Got it. But if the EU asks you—

I mean, we get requests, but the honest truth is: are we actually solving the problem? Just shipping a robot to the military doesn't solve the problem yet. We really need to go deep, so you would need a different team to do that—our team—

Really? You need a different team? Well, it seems like you could use the same team and build military applications.

No, autonomy is very different, right? For example, we do autonomy. You have time to set up a robot, and it does inspections and all of that. In the military, it's about milliseconds. It's remote control, human in the loop, different communications, and different autonomy. Then everything on top—application software—is very different.

Yes, you could use a 4-legged robot to also go into a house.

That's about it, right? The rest is different.

How do you think about robots that are armed? Clearly, China has done demonstrations of these same types of 4-legged robots with guns on them. Obviously, with AI, these Terminator scenarios are here.

Yeah, they're being built in China already.

Yeah. We've seen drones on the battlefield in Ukraine. Norway is not far away from Russia. It's not that close, but it's not that far away either. How do you think about the fact that communist countries are building these robots that have weapons on them?

Yeah. I personally don't like it. I hate it. I think it's concerning. As an engineer, you should have pride in building technology for good. Defense is one part. Actively attacking and putting a gun on it is risky. These technologies are getting mature, but they're not mature enough that you would put somebody else in harm's way.

Yeah, the enemy we're going to be faced with is going to do this, and we need to monitor it. What is the buzz inside the industry about this? When you're out with other people in the industry, what do you know that we don't know about what's happening in those authoritarian countries with robotics and the military?

I think these are all very early tests. If I look at those videos, these are demonstrations.

Got it.

I've not seen these types of robots act as drones. Yes, Ukraine—that came out of necessity. That was a mature category that was used in robotics.

Actually, to the people I speak to—I mean, 4 years ago, we wrote a letter together with our friends at Boston Dynamics and others condemning the weaponization of robots for exactly that reason: as engineers, we don't want to see them being used that way, and we think it's dangerous, risky, and stupid.

If you know 1X, they make NEO. NEO is a household robot. You've sold a lot of pre-orders, and you guaranteed people this would make it and ship in 2026 into their homes. What does it cost, and are you going to hit your self-imposed deadline?

You have to keep your promises. Okay. So we will ship in 2026.

Okay.

Now, expectation management here: it'll be slow in the beginning. We want to do it right. But there will be a handful of customers who get their NEO in 2026, and I'm so excited. I can't wait.

What is the cost of NEO?

That's an interesting one because it depends a bit. When we launched the pre-order, we had 2 different payment models. You had an early-adopter, upfront, full-payment option, and then we had a subscription fee. The product, of course, is going through a lot of development, so how this subscription model will look is still evolving.

We want to figure that out together with our customers in the beginning. Another big thing—we haven't really announced this yet, but I've hinted at it a bit—is that we're going to allow a lot of people to build on NEO. We're also launching NEO as a platform.

Yes, I'm thinking of an app store of sorts, or a skill store. If I have it in my home and I want to make a salad, you, as a hacker, could make a salad skill, and I can buy and subscribe to your salad skill.

Yeah, that will be part of it. But to me, NEO and 1X are about so much more than just consumers. Consumers are an incredibly important market, but 1X has always been about how we create an abundance of labor across society through these humanoids.

I sincerely believe that we have a platform now that is so uniquely capable and so well situated that allowing people to build on it will open up how to use NEO across all of society, not just in homes. It will also benefit consumers because there will be more things developed on NEO.

Part of that will be an app store targeted toward consumers, which we're very excited about. But in general, it will be about creating a bigger ecosystem that can accelerate autonomy and the path to having a fully autonomous agent at home that can do it.

What was the pre-order? 20,000 or something? I'm trying to remember.

We haven't given out official numbers, but it's pretty significant. We sold out the first 10,000 in the first few days.

Oh, so people put a deposit down for that. They'll have the ability to fully—sort of like the Tesla $500 deposit, or $500 a month, $1,000 a month, something in that range?

Yeah, $500 a month.

$500 a month. So this is, if I were to think of a parallel, Google Glass or the Vision Pro. This is for high-end folks who are the vanguard, the earliest of the early adopters.

Yeah, 100%. We tried to be very transparent about this. Getting a home humanoid in 2026 is going to be rough around the edges.

Right. They're going to fall.

They're going to fall, right? But I am very happy to say that I think we will actually be able to ship something that's very close to full autonomy, which we did not want to promise when we launched this because it was too early.

I'm not going to fully promise it yet, but the way it's trending now, it looks like we will be able to ship an experience that is fully autonomous and still quite useful.

Now, if you want everything to just work out of the box on day 1, then there will be some teleoperation involved, or some guidance of the system. But the thing that really excites me these days is that we're seeing the path to actually shipping something that, if you want it, can be a fully autonomous experience, and it's getting pretty darn good.

The teleoperation is fascinating to me. I don't know if you saw this, but in New York there was a chicken sandwich shop that couldn't find a cashier. So they hired somebody in Manila, in the Philippines, for $3 an hour, which is a huge salary for a cashier in the Philippines. They had her on a Zoom call, and she acted as the cashier herself. You could order, and if you had a customer service issue, you just talked to her and she was like, “Hey, I'm right here.”

That is, in some ways, what you'll be able to do with your robot. You'll have somebody in the Philippines whom you'll be able to tap into, who'll be able to turn it on, and when you say, “Hey, pour me a glass of orange juice,” that person will be able to remotely do that task. Is that what I'm envisioning here, correctly or incorrectly?

I think it will all happen. So, back to how the platform works, right? Let me just back up and spend 2 minutes on that. If you think about Neo as a platform, if you want to build your own shop around this orange juice shop, you buy a bunch of Neos, you get the robots, and you get the robot operating system with fleet management and all that.

You also get the data collection equipment, which is gloves that have the same tactile sensors as Neos, and the same vision system. You can gather data in your shop, fine-tune our model within our system—we do all the dense captioning of the data for you—you fine-tune your model, you deploy it, and you get this working. Now you have a fully automated shop, and you're very happy.

That's one path. Maybe that doesn't quite work, so you say, “Ah, I'm going to have someone intervene sometimes in teleop,” and then your data gets better. That's one way of doing it, right? There are many ways of gathering data. Or maybe you're just saying, “You know what? This is super complicated. I just want it fully teleoperated.” That's also fine. It depends on how you want to apply this.

The platform goes all the way from developers who just want to automate their workflow to the more foundational labs that want to deploy their models. There's also a world where you can run someone else's model on Neo. We're going to allow that, I think.

So you're going to be an open platform. You'll be, in a way, agnostic to the knowledge inside of it. You'll be able to plug in—if OpenAI has a world model, or Claude, or some of the other independent world models, they'll be able to be plugged in.

Yeah, 100%. Now, I sincerely believe that our model will be the best one.

Sure.

And I believe in competition. So if we, who actually control everything from the manufacturing all the way up to the product, can't make the best model, then we kind of failed.

Yeah.

Will we allow other people to build on this? 100%. One of the big reasons for this is that, currently, if you look at where the field is, there is no one general model that solves everything for robotics. It's not there yet, right?

If we're stuck in our customers' backyards helping them integrate with ERP solutions and everything else over the next couple of years, we're not going to get there. What we want to do is work on the general problem: How do we solve embodied AGI so we can actually create an abundance of labor?

This requires us to focus on the general problem and then allow other people to also help apply what is available today and help build the ecosystem. If we get this enormous robotics ecosystem, we all benefit.

Yeah. I could see some applications where one teleoperator—let's say this was a convenience-store robot that just helped you carry stuff out to your car. That might only happen once every hour. You could have one teleoperator, or maybe you have 10 of them monitoring 30 or 40 Neos, and they control them remotely and help people move the groceries to their car.

Yeah. Personally, actually, I have a use case for Neo.

Okay.

In teleop, which is—I’m part of the time in Norway, mostly in the San Francisco area now, but part of the time in Norway, and I'm also at conferences like this, right? When I'm traveling, I want to be able to be present and run my company through Neo.

Yes.

Put the hat on Neo: “I am Neo.” That's actually pretty magical. I can go around, pick up the parts, look at the parts, talk to people, and be in the meetings. That's one application of teleoperation that I think will never go away. No matter how good your autonomy is, that will still be there.

Yeah. Your avatar at your factory in Shenzhen.

100%.

Yeah. There are other applications like this, such as remote power stations where there's no one within an hour's drive. You have a robot standing in the closet, and something goes wrong. You go out, flip the old switches, and do the things. You're likely not going to automate that because it's a one-off thing that happens every few months, right?

Right. Essentially, it's what used to be called an “expert in place”—the concept that you can teleport the world's best expert anywhere in the world to help solve a situation.

Like a surgeon.

Yeah. It's super useful. I do think that what we've experienced over the last year is, first of all, that Neo has become so capable, especially with the new hands, that teleoperation does not fully use the hardware. You're not able to get the teleoperation to be good enough to fully utilize the hardware.

So the fidelity of the hand is greater than a teleoperator is able to leverage.

Yes, right. The teleoperator will not feel the same thing that the robot is feeling, for example. Then you need to build full haptic systems, and they're going to slow you down and be slow and clunky.

So we're increasingly seeing that gathering data with humans just wearing the sensors of the robot, in as transparent a manner as possible, is the most useful. They should not be disturbed by what they're doing, right? That's the most useful data to solve baseline dexterity on the robot.

But even more importantly, the big bet that we made—which is this decade-long bet in 1X—is that if you get the robot to be similar enough to a human, then you can train on all of the available video data out there of humans.

Yes.

We're starting to see some very good proof that this is actually working incredibly well. That's the reason we started the 1X World Model Lab, because we now finally have the scaling laws for that. We're seeing that this—

Take us inside that. Take us inside the lab. Are you having people in factories wear glasses and wear your hands and do their tasks over and over again? Are you working with the Micron of the world to do real-world stuff and outsourcing unique proprietary data that you can have that other companies don't? How does the world model get built at scale?

So, first of all, yes, we do that, but that's not the main point. Ultimately, it's very simple: The model is going to be as good as the data.

If you think about the data pyramid, then on the top you have teleoperation data: a very high-quality, small, fine-tuned data set. What we do is have the operator try to do the task very well and very fast. They will often fail, then just try again, and we pick the good samples where they did the task as well as a human would.

You don't need a lot of that data; it's just to align your model. Then you have the data that you're talking about, where you put the sensors on a human and gather data.

Yeah.

You have more of that, and it's very close to the robot, but it's not the robot. Teleop is the robot; this is not the robot, but it's close.

Then you have egocentric video from a human's point of view. That is further away from the robot, but it's still quite close because the robot's hands are the same as human hands, and they look the same.

Then you have general video data.

Yes.

Of the world—or the world in general—and of people, right? Because Neo is so similar to a human, we can actually utilize all of that data.

Now, the bottom layer in the pyramid, which is this general video data, is absolutely ludicrously immense compared to anything else. It's YouTube; it's everything. If you look at what is needed to actually achieve true intelligence, you need multiple orders of magnitude more data than anyone is even close to collecting over the next few years with egocentric data or with sensor data.

Got it.

All of the major breakthroughs that we've seen, as far as I'm aware, in AI have been because someone figured out how to use a huge new data source that previously we were not able to use. You unlock some new set of data, and now your model capability greatly improves.

Well, you've got a lot of people out there trying to find data like that. It's going to take years.

So, it’s like a catch-22. Our big bet is that you have to be able to utilize the general video data out there.

Yeah.

The only way to do that is to care about every single tiny detail of the robot, so it’s as close to a human as possible. Flesh, tissue, and skin are highly nonlinear. How much force does it take for them to deform? What’s the friction? What is the impact energy when touching the table?

Yeah.

People also have different-sized hands.

That’s a really good way of saying it: wingspan. We’ve always called it the gorilla coefficient. Long arms.

Yeah.

Yeah. If you—

But anyway, my point is, yes, we do all of these things, but ultimately what differentiates 1X from all the other robotics companies is that we are all in on pretraining our own models on video data from the internet.

Yes.

Our cross-embodiment is not another robot. Our cross-embodiment is the human. We want to be as close to that as possible because that solves the catch-22.

In the end, all the data will be robotics data because robotic data has the actions, the tactile information, and the forces. It’s better. But the only way to get all of that data is to create a base model that is good enough to deploy all these robots across society, where they will do useful things that people pay for and also gather the data.

When do the robots become recursive in nature, teaching themselves and building themselves? We’re seeing this with large language models now, where people are creating agents. Instead of giving them prompts and instructions, we’re starting to say, “Here are the goals. Here’s a loop. You’re one agent that identifies potential customers for a business. You’re the agent that does customer success, and here’s what that looks like. You’re the agent that does pricing of products.”

Those agents start working in concert. We’re starting to see that in knowledge work. When does that come to robotics, where you don’t have to actually worry about making the robots better? They’re sentient enough—to use a word that’s perhaps not accurate—but they know what their mission is. You’ve given them the goal: “Hey, you’re working in a Michelin-starred restaurant. Your goal is to make the most delightful food with this level of fidelity and perfection. Here are the outcomes.”

And it says, “Okay, I’ve just got to get better at poaching these eggs to really be great at this.”

It’s kind of sci-fi.

No, no, it’s not sci-fi. It’s actually something we think a lot about, but it’s also incredibly hard to answer because the development is going like this, and you’re here on the curve. When you asked me a year ago, I was much more bearish on how far along we would be today with AI. Every time I sample things, they’ve moved faster than I think. It’s easy to get carried away, right?

But if I try to answer it broadly, I am extremely sure that we’re less than a decade away from hard takeoff. By hard takeoff, I mean robots building the robots, the data centers, and the chip fabs; doing the mining and refining; and creating a true abundance of labor—a self-sufficient system.

Under 10 years?

Under 10 years. My current bet would be 3 years.

Got it.

If it takes 10 years, in the history of humanity, it’s still a blip. It doesn’t really matter. That gets back to what 1X is, right? Because—

And you call this the industry term “hard launch” or—

Hard takeoff.

Hard takeoff.

Takeoff, not takeover. We’re going to do it right. It’s going to be hard takeoff.

Hard takeover. Yes.

But I’ve heard the term—this is an industry term—AGI. You can’t really get this without the physical part, right? Digital intelligence can never create its own substrate. You need the physical part.

Right.

I think this is also going to have an incredible impact on humanity with respect to, for example, progressing science. A lot of the demand we’re seeing now on our platform is from people who want to automate lab work. If your AI model can’t actually build and carry out its experiments and observe the results, how is it going to progress science?

All of these things will happen in the coming years as AI becomes physical. The exact timeline is a bit hard to predict, but it’s years, not decades.

Yeah. If you believe it’s 3 years—and I know you’re an optimist. You have to be to do what you’re doing, a crazy optimist for sure—and you think the outer estimate is 10, we’ll be fine with 5, 6, or 7.

Bernt, you’ve got to catch a flight. This is amazing. Continued success. If people want to order a Neo and give you $500 a month to be part of this absolute lunacy that you’re doing, what do they do? How do they get in?

You go to our website and order a Neo.

That’s it. It’s that simple. It’s 2026. It should be that simple.

It kind of should, right? If you can order a Tesla online, you can order a Neo online.

Transparent pricing. I like it. Yeah.

Amanda McMaster is here—not McMasters. McMaster.

Just McMaster. No.

Just McMaster. No McMasters. You’re the interim CEO of Boston Dynamics, the OG, the original robotics company. These are the robots we’ve seen for decades doing backflips, doing kung fu, getting kicked and beaten, and getting back up.

We’ve been having a hard time remembering who owns this company now because it was an independent, venture-backed company. Then Sergey and Larry bought it. It was part of Google, then it got sold. I think Masayoshi owned it at some point, but I believe Hyundai owns it now.

That’s correct.

Did I get that whole history correct?

You did. You nailed it.

Apparently, I read way too much industry news. Now you’re in charge of this.

Yes.

It’s changed hands many times, and you went from being essentially one of one in humanoid robotics to one of many. We’re here at the Machina Summit in Paris, and you see many contemporaries now. What is Boston Dynamics working on now? Is it still a research project, or are you going into the real world and applying these robots? I think you got there early, but you now have to deal with fierce competition.

Yeah.

Yeah, we are big on deploying robots. It’s no longer an AI lab experiment or a research and development company. We’re now focused on real-world deployment.

We started with our Spot robot, which many people know. That’s our mobile quadruped.

Famously in Black Mirror, chasing people down—not yours.

Oh, you can own it, right? There’s always going to be a dystopian version and a utopian version. You’re obviously pursuing the utopian, but that is a really cool robot that has been deployed.

Yes, it has been deployed at real customer sites. It’s providing real customer value. At this point, we have over 500 customers in 46 countries. It is the mobile autonomous robot that’s used more than any other on the planet right now.

Wow. So it is the most deployed and most utilized.

Yes.

What is the number 1 use case for it? Is it security? Is it inspections? What do people use that dog form factor for?

Yes.

Or pony? What do you like to call it? Pony dog.

We like to think of it as a dog. I think it moves like that. The customers are finding a lot of value in industrial inspection. They’re using it for acoustic and gauge reading, as well as vibration detection.

If they have expensive assets in their facility and want to monitor them, this allows them to do that. It can do that during the day and then perform security perimeter work at night. So the answer is yes, we do all of that.

The real inflection point was customer ROI. We want customers to find value in this and have it do really useful work. It’s not just about being cute and dancing; it’s long past dancing at this point. It’s now doing real work. Customers need to see their ROI in under 2 years.

Those inspections, if they were even being done, were being done by humans.

Yes.

Humans, as we all know, are fallible. We make mistakes. These robots are now out there as little puppies running around a water treatment facility, a bridge, or whatever it happens to be— infrastructure, pipelines.

They can record many different types of sensor data: video, obviously, vibrations, radar, I’m assuming, and all the different types of acoustics you mentioned.

What do those robots cost? What’s the range of the hardware cost? And what’s your business model with these? Do people buy them and rent the brain?

What do you think, as the CEO, will be the business model, and what is the business model with these hundreds or dozens of customers deploying hundreds of these?

Yeah, so we went with a CapEx model to start with Spot. We'll probably be doing a robot-as-a-service model, likely with Atlas. We understand that with the humanoid form factor, folks may want to spin them up at different times and have the ability to decrease usage. With Spot, CapEx has been pretty effective. It's the way these industrial customers think about industrial tools, so they generally want to spend CapEx for this.

Depending on their configuration, it ranges anywhere between $100,000 for the base robot all the way up to $300,000 for a one-year, fully loaded deployment with services, integration, and deployment. It's the price of a Tesla to a Ferrari, depending on how you equip it.

But people need to understand that the lifespan of these is greater than 5 years, I would think. These are known for industrial use, so if it can run—I'm assuming you can run 20 hours a day, 22 hours a day with charging.

Yep. We think about it in terms of mean time between interventions, and we're at over 3,000 hours. Only a couple of times a year does a human have to be involved, and it has a charging station. The battery runs for about 90 minutes. Usually, we'd have 2: one comes back, sits down, and charges, and the next one can take over.

Does it automatically swap the batteries, or—

It just sits down onto its charging port.

Perfect.

Yeah. Atlas has swappable batteries, though.

Yes, but the hot swap has to be done by a human.

No. Atlas does it itself.

Oh, it does it?

Atlas will have 2 batteries. It turns its torso around, and you replace one with the other one. It always has a backup, so battery life is not perfect. For the humanoid one, it can do it itself. Obviously, the dog gets charged, so realistically, they could be in the field for close to 24 hours, maybe 18 to 20.

And so that puts the operations at a couple of dollars an hour. Has that changed how people look at the use case—the dramatic lowering of cost? I'm assuming union workers inspecting pipelines are getting paid $40, $50, or $60 an hour fully baked, with their benefits, their pension, and whatever else. It's quite expensive.

We haven't necessarily looked at labor replacement for Spot. While that is a metric you might look at, we thought about how to bring Spots in there to augment human labor. One, humans weren't doing the task; even if they were tasked with it, they weren't actually doing it. And 2, we're just trying to figure out ways that humans can do more knowledge-worker tasks as opposed to going and doing inspection.

So yes, one of the metrics a customer might look at for ROI is labor replacement. We're leaning more into how much we save you. We found an air leak in your facility, and that would have been $3 million a day at one time.

Yeah. The outcomes matter.

Yes. So what is the value that we're driving?

But is it still delicate in the industry to talk about labor replacement? You have to be very thoughtful about that in this moment in time.

And let's be honest. There's going to be an element of labor replacement for this as a metric because it's easy: how many bodies are in the world, and how can you imagine a total addressable market relative to that? I just don't think it's the only conversation we should be having, right? It's just an element of it.

And hopefully we're getting rid of the dangerous jobs and the ones people might find oppressive.

Yeah.

Dull, dirty, dangerous.

Dull, dirty, dangerous.

Yeah. We don't want that hurting their bodies.

Yeah. We only get 1 human body. Yeah.

The Atlas—how do you think about onboard compute versus remote when you put the amount of brains—my understanding is you have the brains on the robot.

Yep.

That means crazy battery drain. What do you think about the option of having the brains in the cloud and having these be more lightweight if they're in an area that has extremely high-speed Wi-Fi, et cetera? Do you offer that yet, or is it all, “Hey, you've got to have a robot with a lot of brains on it because that's what the customers want”? That seems to be a paradigm shift that's occurring now. So how do you grow that, or how should we think about it?

We think about 2 brains, right? It's my simplified version of telling the story. There are 2 brains: the brain that controls the physicality of the robot, which is what Boston Dynamics is known for—the dynamic movement, reliability, and the way it manipulates things in the world. That lives on the robot.

The reasoning layer underneath that gives you the semantic understanding of its environment; that can be in the cloud. That's something that we might partner with Google DeepMind on, or we may partner with other AI partners, or we'll build some of this ourselves.

Then the wrapper around all of that is the very specific information that a particular customer needs around their own workflows: the way they think about the job processes they have, the tools that exist in their facility, and how this robot will interact with them. That's going to live somewhere in between. It could be on the robot if you needed it to, or it could be in the cloud. We'll figure out the wrapper for that.

What percentage of the robots are built in the United States or outside of China and Taiwan today?

100% of the robots.

100%. So there's no issue with the sovereignty of robots in the United States. We're seeing a lot of cheap robots coming out of China.

Yeah.

Your personal opinion as the CEO of this company and as an American: under any circumstances, should we allow humanoid robotics from China in the United States?

No. It's not safe, right? We've already heard about leaks happening with some of the quadrupeds that you're seeing in the United States, and it's being back-channeled back to China. Listen, we have seen what happens if we let China win in the semiconductor space. We can't do that with robotics.

So we need a concerted effort to protect our IP and make sure that we're bringing manufacturing of this ecosystem into the United States or into allied countries. That means that we need to take our national robotics strategy. We're lucky enough that we get to sit at the table in some of these discussions. I'm hoping that more companies in the United States join us in taking up this mission.

Yeah, we have to be pretty serious about this. It's an existential issue. Not only do we have to win this, we have to make sure that the rest of the world uses our platform rather than China's. How do you think about the military application of these? Obviously, the field has been changed with drones in a way and at a velocity—no pun intended—that I don't think anybody anticipated because of what's happened in Ukraine and now in the Middle East with the war with Iran. How do you think about Atlas and Spot on the battlefield? Where are they in terms of deployment in the military?

We've been pretty public about the fact that we have an anti-weaponization stance. But listen, for what we're trying to do right now in industrial use cases, it's a distraction for our business.

So, focus.

It's focused.

It's not philosophical.

I mean, it depends on who you ask. As the CEO, I'm going to look at this and say I'm all about focus right now. We need to be focused on the markets that we think we're going to win in. Certainly, we have great ties with the government, and we're happy to do any non-weaponization work with them. We do that today.

Okay. So you'll have them in—or you do have them in—the field. Maybe if one had to go collect a soldier or bring a med pack, you'd be okay with that. Disarming a bomb, you're okay with that?

EOD—explosive ordnance disposal—is one of the great use cases for robots.

And you're doing that currently?

We do that currently, so we're okay with that. What we don't want is Terminator robots, right?

It's not good for the market, but China's building them. So if China's building them and we don't—right?—you're kind of obligated, if you're Boston Dynamics, to build them. So if China puts these into the field, will you build them to protect America?

I think that's a tough question, and I think we're going to have to answer it when the time comes. Hopefully, it never comes.

The time is going to come. I can assure you.

And I can assure you what your answer will be when President Trump calls. You will say, “Sir, yes, sir,” or else your company will be nationalized. I mean, this is the reality of it. I'm being a little facetious and playful with you, but they're going to deploy these, and they already have shown—

You've seen them put AK-47s on these.

Yes, not on our robots.

Not on yours, on theirs.

Yes. And listen, it's terrifying. Terrifying. I know that we have the best robot and the most capable robot in the world. If and when that time came that we had to make a tough decision, we would make the right one.

But today, we don't have to make that decision. So I'm going to keep everyone focused on the application space that makes a lot of sense for us to make money.

I'm going to tell you a secret.

Don't tell anybody.

The CIA, the FBI, and the Department of War have many of your robots with many weapons attached to them currently.

Don't tell anybody.

All right. Listen, I know you have to go. Continued success. This is such an important American company, and I hope you take the job and become full-time. I know you're interim right now. I wish you great luck with it. If people want to come work at Boston Dynamics, please tell us where you're based.

We're in Waltham, right outside of Boston. But we're open to some remote work, and we're considering coming to the West Coast.

I was about to say, I know it's in the name, Boston Dynamics, but I assume that, with all that talent accumulating in the Bay Area, you're going to need to pop up a space there.

Yeah, we're consistent.

All right. Listen, continued success. Thank you so much.

All right, everybody. Our next guest is Professor Jonathan Hurst. He's the co-founder and chief robotic officer, or chief robot officer, at Agility Robotics. You have a PhD—

In robotics.

From 2008. So you've been at this for over 20 years.

Well over 20 years.

Things seem to have heated up in the last 36 months. Maybe you could, for the audience, before we get into your product line, level-set what you've seen in the past 20 years.

Yeah. I mean, 20 years ago, when we were doing this, it really was unknown in industry, right? Robotics was more about automation systems.

Yeah.

And in the research community, we were doing things like humanoid robots and autonomous mobile robots, really trying to build the intelligence and then build the hardware that can make it capable. That's really started to break through now into the real world, into having direct impact beyond being a research topic. The universities have seen this demand and this growth, and people love robots. There's a lot of demand from students who want to do it, so the number of programs has grown, and it's just exponentially growing. Very, very exciting.

We've had a lot of false starts with humanoid robots, which you're specializing in—

And AI.

And AI. They call it AI winters—multiple ones.

This time is real.

Quite obviously. Explain to the audience why this time is different and why you believe this time we're going to see robotics, and humanoid robotics specifically, deployed at a scale that I think we can both agree will be, maybe in the next 20 or 30 years, one-to-one with humans on the planet.

Very impactful.

Yeah. Why is this time different?

Yeah. Well, I would say, generally, it is very easy to make a robot that looks like a person, and that's why we've seen humanoids for 100 years and more. It's very hard to make a robot that can do useful things in human spaces. We're starting to see that today, and that's the difference. Even if it doesn't look exactly like a human, but is maybe a little bit humanoid, it's doing useful work. That's where the impact matters.

Because of large language models, a lot of things have now become free. When these robots look at a table here—

Yeah.

—and you say, "What's on the table?" It knows that's a phone. It knows this is paper, tea, and water. It probably knows how many ounces are in each.

Yeah.

If it were sitting here 3 or 4 years ago—

It wouldn't actually know—

—what was in the world. You would have to program it in a very narrow way. Perception was incredibly difficult, and the fact that perception is all but solved at this point is a really, really huge inflection point. I said, yes, robots doing useful things, but also people can now see the future of generality. AI is really enabling much broader context awareness for these robots, so people can see that this is going to be useful, generally doing many useful things very soon.

So there's perception: The robot has to understand the world.

Yep. But then there always seemed to be this blocker with getting the robot out of a very confined, narrow task, like in a factory—

And my perception is that it was the communication and the training level. Maybe we can unpack that a bit, because my understanding was that previously you basically had to hard-code the robot if you were going to make a cup of coffee. We have a company I invested in, Café X, and it is a robotic arm.

Mhm.

It makes a cup of coffee perfectly every time, can draft a beer, all that stuff. But it had to be manually coded. Now, the instruction set, because of perception and because of language models having trained on every video on the internet and every coffee recipe, also seems to be free. Am I wrong, or—

Not yet. It's actually quite different. Language models—think of them like they're now becoming kind of a commodity, like the internet. They're available to everybody. It's this amazing rising tide. But these language models are trained off the entire data set on the internet, and that data does not exist for robot control. What's the example for your robot of all the torques, all the torque commands to every motor, given all the sensor input? There's no training set of data, so you have to generate and create that somehow.

There are a lot of different approaches and ways people are going about this. Some of these AI tools—again, think of AI not as a black box, but as a big tent of many different, very different, useful computational tools. In order to control a robot, you can do these things by learning from demonstration. You can teleoperate the robot to train from that data. You can give it animation input or motion-capture input, or any number of different things.

But that's also got a real hard limit, because a person controlling a robot is not really getting to what the robot could do if it were optimal and how its behavior could work. A robot needs to practice.

And that's where you get into world models and sim-to-real transfer and all of these kinds of things.

And world models are the next frontier. People are literally putting gloves on humans and having them control robots remotely to actually chop and make a salad, to pour water, and that's being done today by many different companies. World models will solve this problem—

Or—

—they are part of the solution. As with all of these things, there is no silver bullet, right?

So the world models, as I understand it, are: Can you model an entire warehouse and all of the physics of all of the objects inside of it, so that simulations of these robots can go practice in the world model without breaking things in the real world? You can compress it so you can do 1 million iterations within days, computationally, and things like that. But there's always a massive sim-to-real gap. Things aren't simulated perfectly. As you pick up something in the real world, there are wave dynamics and condensation on the glass, and the dynamics of the robot are not perfectly modeled. All these things are still very, very difficult. That takes real practice in real life with robots in order to—

Yeah.

So, is there going to be a singularity or a crossing-over moment where recursive learning—just putting the robot in the kitchen, letting it make its own mistakes, and then saying, "Do the next test, do the next test"—which is how we taught it how to win at chess or Go? We didn't tell it, "Here's how to castle." We just brute-forced it and said, "Try every computation," and it was able to figure it out. Now, with these recursive loops, what will get us there quicker? Somebody builds a world model, says "Go recursive," puts the robots into a kitchen, and breaks a lot of china? Or is it going to be these world-model companies very refinedly working with a human alongside a robot in a Michelin-starred kitchen to make that soufflé?

I mean, that's not a very satisfying answer, maybe, but it's all of the tools. All of them, right? There's not a silver bullet at all here. I don't believe that there's this singularity. I do believe that things are going to get better and better. Think of it more like a snowball picking up steam going down a hill.

Got it.

But the reason that it's snowballing like this is because people are putting money and resources and engineering time and engineering effort in as they explore everything and start to figure all of this stuff out.

All right. So—

But humans, for example, have evolved to learn. We are very good at learning, and it takes very little data to show us how to do something. Then we practice and practice and iterate. Robots are not very good at learning yet. Robots take so much more data, so many more examples than a person. We're still figuring out how to teach robots how to learn.

But one of the benefits that robots have in the long run is that they've got Wi-Fi. When you learn how to play the violin, you can't just load that onto somebody else and have them learn how to play the violin based on your learnings.

One robot learns to play the violin, and all robots know how to play the violin.

Or all robots of that type know how to play the violin.

Right. Yes. And then minor variations for the next type and the next piece of hardware. So you're actually deploying your product—it’s called Digit. Digit is, I think, 4.0. You're going to release 5.0. You've got, let's say, dozens in different applications out there in the real world. Give us an idea of what the forward deployment looks like today and where you think it will be in a year or two.

So today, it's doing these multipurpose workflows that are still reasonably well scoped, like picking up bins and totes and carrying them around. The reason we do that is because you need 2 arms to pick up big things. You need this whole-body control to be dexterous in how you're manipulating and moving those. You need to be balancing to lift them to the top of a tall shelf in a narrow space, so it kind of justifies the form factor for this one use case.

But the really useful aspect of a humanoid is its versatility. When we do the each picking, fill a bin and carry it somewhere, palletizing and depalletizing, and expand out into more and more use cases, that's when it really starts to escalate. Digit V5, which is coming out later this year, is the first time that a humanoid robot—a robot that is balancing—can step out of a work cell and does not need a physical barrier between the robot and the person to maintain safety in this warehouse. So when Digit V5 is out there, that's kind of the scaling moment for us.

Yeah, this is a key moment that maybe people don't appreciate. But if you've ever been to one of Elon's factories or Toyota's factories, there are lines. There's a line, and if you cross that line, everything shuts down. I've taken many of these tours with Elon, and they're like, “Seriously, please don't cross that line, because it's going to cost $1 million if you do at the Tesla factory, because it's cranking.” We're starting to feel comfortable enough that these robots aren't going to fall over and break somebody's ankle.

Well, it's been a very, very intentional process over the past 2 or 3 years. This was our experience with Amazon when we deployed: The robots are doing the task, and they're like, “Great, it solves all the R&D goals we had.” We're like, “Great, let's go deploy.” And they're like, “Oh, no, no, we can't deploy,” because they don't meet our safety requirements. It's like, “Okay, how do we meet that?” Well, it turns out that's super hard.

It's been a bottom-to-top design of this machine. Holistically, every system of the robot is touched to figure out how to make it safe.

When we look at an industrial-strength robot like yours, the bill of materials—

Uh-huh.

Tens of thousands of dollars each?

Yeah. I mean, we're not discussing bills of materials. We know that the costs are coming down and down and down over time. We'll be selling robots in the vicinity of the cost of cars and things like that. The real question to ask is: What is the value that they produce? When you have a robot that's working 24 hours a day and has a 5-year life, what's the value? It's quite a lot.

Yeah, it would be. If we were to think about it from first principles, they can reasonably run 20 to 22 hours a day, and then they have to charge.

That's right. So we take 20 hours a day—

24/7, by the way—20 out of 24 hours for our Digit V5 robot, because of the very fast-charging technology that's gone into this battery.

Yeah. So we have 20 hours, 365 days a year. Now you're at 7,300 hours a year. Let's put it at 8,000. Over 5 years, that's 40,000 hours of work.

It adds up.

Yeah. People tend to think these things are going to cost $20,000, $30,000, or $40,000.

They will at some point.

Yeah.

It's going to need to go through the scaling process and have 100,000 robots out there before that actually is real.

So that's $1 an hour. These people are being paid $40 an hour in factories currently.

Yeah.

Maybe in some other countries, $10 an hour, but let's put it at $20 an hour. You've got 90% cost compression at some point when these things hit the market, which gives you plenty of room to charge Amazon or Toyota, or other partners, on an hourly basis. Is that the current plan—to charge per hour of utilization? You own the robot. They—

We do both. We do a CapEx model for customers that prefer that. We also do robot as a service for customers that prefer that. It's a lower barrier to entry and lower risk for them.

What's the price of a robot per hour?

We're not talking about that right now. But I will say that as the robots get better and better and better at what they do, their value goes up and up and up.

And that's at the same time that the costs to build the robot are going down. The value for these robots is really set by human labor and what it costs to pay people to do these jobs. So it's a very inelastic price for a very long time.

So between a bill of materials in the tens of thousands of dollars and people in factories getting paid $20, $30, or $40 per hour in the Western Hemisphere, in the modern world—

Which is a pretty big market.

Yeah, pretty big market. Plenty of room for you to save them money and for you to make enough profit—

Build an actual business, you know.

To build an actual business. Yeah. So let's take the conversation to what you think the time frame is. If I were to ask you about Amazon factories—or, if we want to take Amazon out because they're a partner and I don't want to get you in trouble—an Amazon- or Target-like company, at what point will the majority of workers in a factory be robotic? When will that flip happen to 51%, knowing what you know, Jonathan?

I mean, already in a lot of these applications, the majority of the workers are robots.

Sure.

Right. There are a lot of AMRs, a lot of conveyor belts, and a lot of industrial robot arms, and that's not changing. That's continuing to grow.

Sure.

And this is just a new form of automation, like all of the others, that's helping to increase and build that productivity.

So, how do we—in the United States, anyway—build our GDP? It's not a growing population. No—

It's increased efficiency and capability. And the only way we could do that is with more and more—

Especially not with the anti-immigration vibes we have in the country right now, or even in the Western Hemisphere. Let me phrase the question another way. If there were 1 million people working in factories sorting packages, at what point does it go down to 500,000? Is that 3, 4, or 5 years?

I think we've already done that, right? And with these new systems, it's going to just continue. Someday, there's going to be an autonomous truck that drives up to a completely lights-out, autonomous package-sortation factory, and then an autonomous truck leaving again.

At that point, it's probably specialty automation doing those things, because it's just doing them 24/7. A humanoid doesn't make sense; it's not the most efficient thing for that specific task. A humanoid is useful for walking into human environments and doing human workflows. So by the time this one factory is entirely automated, there are also a whole bunch of other factories that are still legacy and still need automation where humans were.

But we're also working now in retail and grocery stores, hospitals, construction sites, and delivering packages to your front door, which is a forever-human environment, right? Front yards and that kind of thing.

That's going to be an interesting one.

Yeah.

Because it's fairly obvious to anybody who has even looked at the latest generation of humanoid robots that the factories are going lights-out. Most people are incapable at this point of imagining a Waymo robotaxi, an Uber self-driving car, and a robot getting out.

Yeah.

And bringing the packages to your doorstep.

That's going to happen.

Absolutely. Going back, are you working with folks on that? You don't have to say who, but—

That was one of our very first use cases that we explored with Ford. There's a nice video online of our very first Digit robot getting out of a vehicle, walking up to someone's front porch, and dropping a package there—stairs and everything. We could do that. This was 7 years ago, something like that.

But I don't think it's the best first use case or the best first market. So it's on our roadmap for sure.

But it's such a big market for deploying what we're doing right now. We're going to start there. When you look at applications, knowing what you know over 2 or 3 decades, what do you think is a use case or 2 that are nonobvious but would be incredibly world-positive?

I don't know what's not obvious. Just picking up stuff and putting it somewhere else is such a huge use case that frees people from the classic 3 Ds of robotics: the dull, dirty, dangerous kind of stuff.

Dull, dirty, and dangerous.

The 3 Ds of robotics. I really hope that our children look back on us the way we look back on some of the jobs people were doing in the 1900s, like coal mining, and say, “I can't believe people did that work.” The number of roles and things that people do today are so much better.

The quality of life is so much better. The jobs that people have today, which you couldn't have imagined in 1900, are often just so much better. I think that's how the future is going to look for us.

You're still a professor of robotics?

Yes.

You have hundreds of people in this graduate program—over 100?

Yes, we do.

Mm-hmm. For young people who are listening to this and are worried about their future and careers, this seems like an incredible career path.

It's a massive opportunity. We live in a time of change. Anytime there's a time of change like this, students coming out have an advantage because all the people who have this 20- or 30-year career and know how things were done have to learn the way things are coming up now, too.

So students have an advantage, and it's hard to predict exactly what careers are going to look like in 10 years. But if students build some of the core skill sets around engineering, those skills are going to be applicable and useful. So there's the PhD/master's version of robotics. Is there another version that's, let's say, a little more general, tool-belt, blue-collar—the equivalent of being an electrician, working on HVAC, or being a carpenter or a contractor?

Yes, absolutely.

What is that, and what will that be?

Robot operators assembling and building robots. The robots can't assemble themselves yet. So there's a lot of manufacturing and, again, robot operations and deployments.

Maintenance—“clanker” maintenance.

Absolutely. Is “clanker” a derogatory term?

I don't know. It's a Disney trademark term.

Oh, is it really?

Probably. Final question. I think we're of the same Gen X. You know General Grievous, the Star Wars character?

You trained in the Jedi dark arts by Count Dooku—

Right? Able to wield 3, 4, or 6 lightsabers at a time.

This is a half-serious question: Why not have 4 or 6 arms facing all directions?

It's a good question. As we think about the first principles of how to make the simplest possible robot to do the task, one arm is not quite enough to pick up big things. You can only pick up small things. With 2 arms, you can pick up big things. Adding a 3rd arm, it's hard to see enough utility to make it worth fitting it in. And then you go to 4 or 5. There's a lot to coordinate and a lot of extra complexity. But what else does it make you do? I don't know. Maybe we'll see that, but it's going to have to be driven by a real need.

All right. What's your favorite robot in science-fiction history?

Probably WALL-E and EVE. I love that vision of these robots just continuing to try and build and create and do what they were designed to do.

I love Baymax, too. Baymax is pretty fantastic.

Wait, wait. Who's Baymax?

Baymax from—what is it?—San Fransokyo.

Oh, yes, of course. I do know this robot, which is very clearly there to help. I love how they show that it does what it's programmed to do. At one point, they remove all its memory, and it turns red and becomes dangerous. Well, that's very real. Your software has to have the safeguards in place. You've got to have the E-stop on these things.

So you think about the prime directives.

Yeah. Basically, yeah.

How do you make sure these things go through the industrial safety process to make sure there's a supervisory circuit, an E-stop on every robot—all of these things that make it so the robots could really never harm a human? Jonathan, I know you're hiring. Agility Robotics is the company. If people are looking for a gig—

It's a fun place to work. Agility is great, and we have a location in Salem, Oregon, where we started, where I am. We have a new facility we're opening in Fremont, California, which is just a beautiful place. That's where we're doing a lot of robot behavior development. There will be robots working all day long, and you can come in and work on them. We have a Pittsburgh location as well.

Oh, right. Right by Carnegie Mellon.

Amazing. Yeah, 3 great centers. So if you're a young person or you're in the robotics field, it's a pretty great place to work. And if you're worried a little bit about your future, go get a PhD or a master's in robotics. Skate to where the puck is going, folks.

Right. Great to have met you, and thank you for sharing all your knowledge.

Thank you.