Brian Heater: Hello, and welcome to Automated. I'm Brian Heater, the managing editor at the Association for Advancing Automation. Very excited to bring you this conversation with Aya Durbin, the director of product for Atlas. This was recorded at Boston Dynamics HQ. Aya was previously at Six River Systems and has a long history of productizing industrial robots, and I think that is very much reflected in this pragmatic conversation about what it will ultimately take to get humanoids out into the world.

If you're enjoying the show, don't forget to like and subscribe. Check out our newsletter over at automated.fm. And with that, here is Aya Durbin of Boston Dynamics.

Brian Heater: You know, we talk a lot about what's coming up next in automation on this show, but if you really want to see the future in motion, you've got to be there in person.

Automate 2026 is where the world's leading innovators, builders, and dreamers come together to show you what's possible. Robots, AI, machine vision, motion control - you name it, all automation under one roof. And as part of Automate this year, the Humanoid Robot Forum brings together leaders, engineers, and researchers for a two-day deep dive into the real-world development, deployment, and commercialization of humanoid robotics.

Register for free at automateshow.com to join us in Chicago June 22nd through the 25th. We will see you there.

[00:02:40] Brian Heater: You're a pragmatist who has been converted into a dreamer.

[00:02:45] Aya Durbin: I am. I would say I'm a dreamer that still is a pragmatist at heart. But working at Boston Dynamics has definitely proven to me that the dreams can become valuable customer solutions.

I'm a product person. I care about delivering real value to customers that provides positive ROI, because I don't believe that humanoids will ever become ubiquitous in society if we can't prove that. The way in which I've become a dreamer is seeing here at Boston Dynamics what it looks like to transform a whole robotic solution to an AI-based approach, and to have the team so focused on proving that that AI-based approach can actually provide value to customers. It's made me think more outside the box of how we take a fully AI-based system and translate that into something that customers can train, instruct, and get to do what they want it to do in their workforce.

Humanoids have to be a trusted and reliable member of the workforce, and the team here has made me a dreamer because they've proven to me over and over again that it's possible to do that in new ways that we've never seen before in robotics. It's been a very fun few years here to see that kind of come to life.

[00:03:57] Brian Heater: We're not going to beat the dead horse of legs versus wheels, because that's been spoken about a lot. People can listen to you talk about that on one of the webinars that you did. But I think that is an interesting distinction here, because you do have to be a bit of a dreamer in order to be on team leg, because that is much more of a long-term solution. If the goal is to get to ROI and get these things out in the field as quickly as possible, you're going to put wheels on them.

[00:04:33] Aya Durbin: I think my answer might be different if I was somewhere else. But at Boston Dynamics, this company has been focused on mobility for 30 years. Walking in the front door on my first day at this company three years ago, what I noticed was how incredibly these robots moved.

I worked at a company that had a wheel-based robot, and the way that Spot moves around the world - and we have a wheel-based robot here - is like nothing you've ever seen before if you haven't seen it in person. It moves about the world reliably. It can walk up and down stairs, and it can walk on graded floors that at my previous company we couldn't imagine working on. I think walking in the front door here at Boston Dynamics three years ago made me realize again that we really can move about the world differently with legs than you can with wheels on a robot. And we've done it with Spot. We've been working on humanoids for years. To me, Atlas having legs is a no-brainer for us at Boston Dynamics because that's something we've already been working on for years and years. So we do have the bandwidth to focus more on manipulation, which I think is the hardest challenge in bringing humanoids into the world, and don't have to spend as much time on the leg part of the problem.

[00:05:49] Brian Heater: You're somebody who's coming from industrial robotics. You were at 6 River Systems and Shopify before this, so you are more familiar with the more traditional systems. You were saying that your family finally believes that you're in robotics because you have something that looks like a robot. But if it's a purely pragmatic goal of getting this into the market as quickly as possible, then it's going to be a fixed, repetitive system. With Atlas, you're focused on the long tail things at the same time.

[00:06:34] Aya Durbin: Yeah, which is not something that traditional robotics has ever really been able to do successfully. But in a world where AI for robotics actually does rapidly expand our capability set for the robot, you have to be thinking about both things. You have to be building out - you need to have a research team that's building out the models that enable you to be a generalist robot, that enable your customers to easily train the robot to do any task in their facility, to be able to easily re-task a robot to do any job in a facility easily. But you can't lose sight of the fact that you need to get that robot into a facility soon to start doing real work, which is why we have been focused on sequencing and logistics tasks that are semi-repetitive, but they're not just pick and place to a conveyor over and over again. We're not doing tote movement. We're doing something where we're working with many variants of car parts or many variants of an item in a warehouse, and picking and placing those parts based on the instructions of an external customer system, dealing with exceptions like no inventory, dealing with problems like where there's no barcode on a part.

All of those nuanced, smaller problems that are the questions customers are asking us how we're going to solve - because they've seen us solve it before with Spot, they've seen us solve it before with Stretch - that's something that's important for us to also be working on right now with Atlas, even though it's not the sexiest research problem. They're problems that have been solved before, but you have to work on the application part of the problem in addition to the long tail research generalist part of the problem, and the goal is that they come together. The goal is that we build one platform, one system, one set of foundational tools that enable this long-term vision. You start with that first application on top of a foundation that enables you to quickly expand within and beyond that application.

[00:08:27] Brian Heater: Spot was obviously a big learning for the company in a lot of ways - the first commercial product after 20 to 30 years. One of the big learnings is definitely that maybe we can't quite do an iPhone model here. Maybe we can't expect the customers to do all of the work of programming all of the skills. And then the other is that we really need to enter the market focused with a single, or at least a few, tasks that we can do really, really well.

[00:09:03] Aya Durbin: Yep. That's definitely something we learned on Spot. We still have customers who use our exposed APIs on Spot and build their own applications, their own functionality on top of Spot. But for the most part, the people that buy Spot buy it for industrial inspection to make sure that their equipment doesn't fail. They buy it for safety use cases - nuclear decommissioning and bomb disposal - or they buy it for academia and build on top of it.

There are other use cases for Spot. People are using it more and more for security, for construction. But when we move into those use cases, when we move into security use cases for Spot or into more construction use cases for Spot, we do it intentionally with partners that are good at understanding what they need out of a product to make it successful, so that when we deploy into that space, it's something we can copy and paste and bring to the next customer with a product that's a little bit better.

When you're bringing a product into a new market to do a new task, you have to start somewhere, give them a minimum thing that we think will work, and then learn from them and make it better - and that's going to be true for humanoids too. The goal is that you can build as many tools in AI as possible to give the customer the ability to build their own features and functionality.

But as we've learned from Spot, and as we've learned across the robotics business, anyone who's deployed any industrial technology knows you need to be thoughtful about showing improvement in your product in order for it to become a trusted and reliable member of your workforce. Starting in target markets on target applications with customers that have safety teams, that have security and IT teams already in place that we can learn directly from, gives us the best chance of making our product great as quickly as possible for customers so that it does become easier to copy and paste and redeploy, and also so that they can help us build those tools and make it easier and easier for them to build their own stuff or to retask the robot.

[00:10:58] Brian Heater: This is something I've been talking to a lot of people about, and I don't think anybody has a great answer - there's a data gap. There's not enough real-world data. I know that Boston Dynamics is doing a lot of different things: tele-op, real-world simulation, video, all of these different things. But the big question right now is, once we start these early deployments and we're trying to get the flywheel of data going, is there a way to incentivize those companies to be a part of that data collection? Is there something in it for the customers themselves to be an early customer, a pilot, one of those partners who is helping sort of develop the system at an early stage?

[00:11:48] Aya Durbin: For us, we're being intentional about who we work with early on, and those customers that we work with are directly guiding our roadmap. I've been in product my whole career, and it's not typical that when you bring a robot on site, that naturally through teaching the robot how to do work, you're guiding the roadmap, guiding the robot capability. The customers we choose to work with are going to be the types of tasks that the robot is able to be really good at in the early days. And so that in itself is an incentive to work with us early on. It's not just that we're saying your feedback will guide us - it will. But also, the tasks we teach a robot to do first are going to be the tasks we have the most hours of training on, so the robot's going to be more performant, more reliable, faster, and able to provide ROI faster.

For us, we built a robot that can lift up to 110 pounds, can lift up to 66 pounds continuously, and work in hot environments. The tasks we're looking at doing early on are not tasks that are easy to hire for. They have high turnover. They cannot find enough people to work in these jobs. And so the incentive for companies is: I can't find people. And it's not just companies - it's operations managers on the floor that are saying, 'I can't find someone to do this because it's backbreaking work. It's terrible. If I find someone, it's for two weeks, and then they quit.' And so there's also incentive there: I need a solution, and I need a solution now because no one wants to do this task, and I can't find another way to automate it.

[00:13:25] Brian Heater: I'm not suggesting that the production Atlas is over-engineered because there have been over-engineered robots in Boston Dynamics history - Handle is over-engineered, and that's why we got Stretch. But it's not over-engineered; it's almost over-capable, certainly for this first set of things it's going to do on the floor. It seems like from a hardware perspective it's effectively designed to be a general-purpose robot, and in terms of applications and AI and capabilities, those things have to kind of catch up to that.

[00:14:07] Aya Durbin: The robot design and the specs that we targeted for this robot were designed to actually have the robot do hard work. So yes, it's designed to do a wide variety of tasks. The height of our robot, for example, is not required for every single task. It's really tall - and that's because if you go work in a warehouse, I don't want to have to walk up a two-step ladder to go reach the top shelf. I want the robot to just be able to reach the top shelf. So we designed the robot to be able to just easily move about the world that exists.

In terms of its capabilities to lift heavy and move differently than a human moves, those specs were designed to actually have a robot that can do the hardest work in industry, because that's where the biggest challenge is for industrial customers, and that's where the biggest challenge is for people. We're looking to create robots that make our lives better, so we wanted to design a robot that could actually do it. We aren't targeting doing lightweight tote movement. We're targeting moving the totes in the building that are 50, 60 pounds, that you're supposed to be using a lift assist to move but no one's using a lift assist to move them.

We want you to remove your lift assist. We want you to save money and space by removing those lift assists. For us, it wasn't over-engineered because we were trying to engineer a robot that could actually do the hardest work and walk into an environment and just work. I didn't want a customer calling me saying, 'Okay, I want to move it from the inbound process in my warehouse to go do picking,' and then when you go move it to do picking, it can't reach the top shelf, or it can't reach into the back of a pallet because its arms aren't long enough. So we designed a robot that met all of those specs, and that's ultimately why you land with something that is as tall as it is and has the reach capacity that it has. It just makes life easier for the robot to be able to do the same things that humans do.

[00:16:05] Brian Heater: I think there's always going to be an inevitable bit of confusion as far as messaging goes, especially with something like a humanoid. You talked about this in that earlier conversation about the expectations that we have going into it because it looks like us. But then that's further clouded by Moravec's Paradox - what the robot can and can't do versus people. Videos of an engineer pulling off a stunt and the robot doing a backflip. It's not engineered to do a backflip, but it can do a backflip.

[00:16:42] Aya Durbin: Yep, and that doesn't mean we made the robot higher cost. We didn't specifically design it to be something that can do a backflip, but the team said, 'Hey, what if it could do a backflip? Wouldn't that be so cool?' And there is value in showing that the actuators are incredibly strong, or look at the sense of balance this robot has. The backflip is a specific example. Even the round-off back handspring that we did in the lab recently - that video is using reinforcement learning to train the robot. A lot of the videos we show seem like they are just playing, but we're actually using the core fundamental technology that we use to train the robot in general.

When customers come in and see the office and we show them some of the gymnastics moves we do, we show them some of the live industrial work, then we bring them back and explain: here's some of the behind the scenes of the videos we showed. They're very fun, but 99.9% of the time, the fundamental technology they're using to do these things is the fundamental technology that we use to unload roof racks, or unload dishes from a dishwasher, or do any other industrial task.

When I used to see the robots doing backflips, when I used to see the robots doing round-off back handsprings or doing parkour, I'd be horrified as a product person. Some engineer is going out there and working on something that isn't core to the customer need. So I'd run downstairs and say, 'Why the heck are we doing this?' And Alberto - the head of our behavior team and one of my favorite people at this company - would explain: this is how this is working. And that really starts to transform the way you think about work in general. You can learn and do science and improve your product in more ways in this new world of AI than just doing the same thing over and over again.

When it comes to AI specifically, the more we train the robot to do all sorts of different tasks, the more capable the robot gets over time. It's been fun to be at Boston Dynamics and learn what the behind the scenes actually looks like and see from the engineering team how we can improve the product in ways that are fun and exciting to show the world.

[00:19:10] Brian Heater: You just completely shifted the dynamic of the conversation, because that is fascinating to me - seeing it through your perspective as a more traditional product manager coming in and seeing these things. Obviously Boston Dynamics - the reason why everybody knows Boston Dynamics is because they got really good at making viral videos, really good at YouTube videos. They became one of the best-known robot companies in the world. They did the Super Bowl ad, all these things. And it sounds like prior to joining the company, you thought maybe these are just ephemera, just fluff, and then you came in and realized that maybe it's more essential to the actual core of the company.

[00:20:00] Aya Durbin: Totally. It's essential to our research process. It's essential to the team morale in general. We spend a lot of time building core foundational technology that enables customer value, and taking that technology to learn how to quickly work with all new types of automotive parts or learn how to more dexterously manipulate something, or test whether or not we can pick up something heavy - maybe we pick up a fun object instead of just a 100-pound weight.

This company didn't get good at making viral videos. This company is full of incredibly passionate people who express their passion through robotics - express their passion through staying on a Friday at 7:00 PM and testing out a new thing with technology they just built. The marketing team got really good at saying, 'Hey, please don't do cool stuff unless you've pressed record on the camera.' A lot of those videos, even in the early days, were someone who was instructed to have at least an iPhone out and be recording when that cool thing happened, and then we just shared people's passions with the world. I think that's why it's resonated so much - you're just seeing people's passion come to life in this robot, which is the best part about working here.

[00:21:19] Brian Heater: We got the tour earlier, and Nick and I were talking about the last time I was at the Boston Dynamics office - it was a smaller one. When you walk around Boston Dynamics, there are a lot of these tweaked-out Spots, like customized Spots, and at the time the big one was the margarita Spot. It was a party Spot. It had a blender on the back, obviously really fun. But what I didn't realize until talking today is that there was a pragmatic reason for it, and it had something to do with the voltage on the back - the voltage in order to crush ice. And it's similar to what you were saying before about how just the way that this Atlas gets up and down is probably because somebody thought it would be cool to do that, and then it served an actual real function.

[00:22:10] Aya Durbin: Absolutely. When you're in this early stage of product development, when you look at something like Atlas, there are so many core things we're testing and core capabilities we're trying to build, and there are so many ways to build those core capabilities. If we're working on high-dexterity tasks, we can pick up something like a screw. I don't care if you're testing with something that looks like a screw but is maybe more fun than a screw, if you can also show that we can pick up a screw.

I love the testing-the-voltage-on-the-back-of-Spot example because you're solving a problem, and the way you solve the problem, I don't care - have as much fun as you want, as long as the problem gets solved at the end of the day, and that's what creates such a really cool culture here.

[00:22:58] Brian Heater: Part of the reason we got the tour before we talked to you is to recognize the scope of what's going on, and that all of these things can be true at once - that you can be working on things that are a little further out there and these very pragmatic problems at the same time, because there are a lot of different teams working on a lot of different problems.

[00:23:28] Aya Durbin: Yeah, and that's what my job is - to figure out how we take that long-term vision and boil it down into the set of tools and products that we need to build in order to bring that long-term vision to life. The goal is to start building that set of tools now and build all of our applications on top of that core set of tools so that we can more easily expand over time.

It's a very different type of product development. I think robotics in general is changing a lot. Product development does not look the same as it used to. You used to not be able to focus on more than one application at once. My pragmatic brain would have been fried years ago trying to think about working on something like sequencing and a task like machine tending at the same time. You just couldn't do both. Now you have to be doing both, and you have to be building tools that enable you to do both. Otherwise, you're not building the product that customers need or that customers expect.

It's been fun to push into that dreamer world and think about how you build tools and capabilities for us and for customers to be able to build one system that unlocks a really large potential for the robot. You're focusing on many things at once, but the goal is actually to focus on one set of tools and capabilities that unlock many opportunities - and that's kind of the way we focus at Boston Dynamics.

[00:24:53] Brian Heater: I think it's a two-way street because the way I've seen things unfold as a member of the media is all of these humanoids are coming out, and it's like: hey, general purpose robots, AGI, these things are eventually going to be able to do everything. And then things settle down a little bit, they're out in the world, and then we start having serious conversations - okay, the reality of these pilots when they're out there is that we're really going to have to focus on smallish tasks at once and then get that data and start to build up from there.

[00:25:29] Aya Durbin: I think that's the best way to approach it, not necessarily because you can't do everything at once - you could try. But as a customer, put yourself in the shoes of any warehouse operations manager or any logistics operations manager. You need to make sure your work gets done and your throughput gets out the door, and that includes the robot being able to handle all sorts of problems that come up for your associates today and for the robot to handle all the problems that it's going to inevitably cause because you're using technology to do this thing instead of a person.

Those niche problems that need to be solved in order to deliver a valuable customer solution need to be addressed. And if you're trying to hear that feedback from tons of customers in different markets and industries all at once - even if you're in the home use case - every single individual person in a home that's going to use a robot has a different perspective on how that robot should perform work, behave, deal with exceptions, deal with problems.

The value for us of working in industrial settings is we can really listen to a core set of customers that are going to give us the best feedback on how to deal with problems appropriately, the way they expect, and be repeatable and reliable, and then replicate that elsewhere. That's one of the things that's exciting about working in industrial environments - just that kind of streamlined feedback that allows us to build the tools to deal with all the problems that come up every day, and build the tools to make sure the robot performs exactly the way you want it to, when you want it to, how you want it to in the environment.

[00:27:12] Brian Heater: Obviously one of, if not the big selling point of the form factor is that it's brownfield - that theoretically you can slot it in. A big, in-hindsight obvious lesson that I've learned since starting at A3 is that people who run factories really don't like if you have to stop the line for any reason. You were at Six River for a while, and I suspect that even when you had a mature system, it was still really difficult to convince people to integrate that in, let alone this entirely new form factor.

[00:27:47] Aya Durbin: Yeah. There's a lot of unsexy things about integrating automation into any environment. Someone needs to instruct a robot to do work. There's a whole IT team that needs to build out new integration points to talk to the robot. That's months and months of work for a customer. It's not impossible to build, but it's hard work for the customer, and it's something that we need to build easy solutions for so customers can deploy with us quickly.

If you want to unlock massive scale for humanoids, you need to provide customers with a new way of solving this really hard problem that all traditional automation has to solve, because it is a big investment to introduce automation. What's cool about humanoids is the potential for them to make an impact on businesses is bigger than anything we've ever seen in the robotics space. The hope is that the investment that companies make is a one-time investment that can be replicated over and over again, so they don't need to keep doing all of the upfront integration, IT, and security over and over again.

It's a challenging process, not just to build the infrastructure, but to get your operations teams used to working with the robot, to get them used to the interfaces they use to communicate with it. They can't stand downtime. When you first deploy a robot, you are going to have to work through some issues, especially if you're working with humanoids in the early years of our deployment process. We're working with customers to understand this is a process.

[00:29:25] Brian Heater: Swappable batteries - that is huge.

[00:29:27] Aya Durbin: They are amazing, and that was intentional. We want our robots to be up and available as much as possible, and we've learned a lot about serviceability and manufacturability from Spot - designing this robot so that it has quick, easy-to-replace limbs, so that if something does happen, you can really easily fix the robot in a triage lane.

We've really reduced the complexity in the robot and made it easy to service in the field to help make sure that customers don't have downtime. We've made sure that our batteries are swappable in less than five minutes so that customers have the robots when they need them. And they're swappable by the robot autonomously. The robots run for four hours, and then the robot can swap within five minutes all on its own. And then there are all sorts of creative things you can do if you need even more uptime than that.

Robots should be able to help each other and jump in and cover work if there is a problem. We've really intentionally designed the system to be there when our customers need it most, and that's another part of the design of this robot that we thought through quite a bit before we started building.

[00:30:43] Brian Heater: Okay, so let's get really unsexy. You've got a great owner and partner in Hyundai in that a huge part of the deployment is going to be them rolling out in their factories. But you move beyond that, and then you start talking about system integrators. What role does Boston Dynamics play in that integration? How do these things actually get rolled out and incorporated and customized into specific settings for specific jobs?

[00:31:10] Aya Durbin: That's a great question. It will depend on the industry and the application. For us in the first few years, we have target industries we want to work in - automotive manufacturing, food and beverage manufacturing, semiconductor manufacturing, and warehouses. In those markets with those customers, the deployment process will be similar to what it looks like for Spot. You go through a sales process. You work with a solutions designer that makes sure the solution you're going to get in your facility is what you expect, that it performs the way you want it to, and that you're really going to get ROI from that system before we even show up on site.

[00:31:51] Brian Heater: But you are ultimately pretty hands-on when it comes to making sure that everything works.

[00:31:57] Aya Durbin: Yes. Especially for our autonomous applications - if we have a Spot going out there to do industrial inspection, or if we have Stretch going out to do truck unloading, someone is going to come make sure that what we deliver to you is what you expected and that it works the way you expect it to.

Over time, there are so many ways we could make this process more customer-based. If customers want to build their own tools and deploy their own robots, absolutely in the future that's something we could enable. In the early days of deployment, we want to make sure that customers get the exact value they expect out of the system, get the ROI they expect, get the throughput and performance and reliability they need, so that we can make sure the business itself is sustainable.

We want to make sure customers love the product, and going on site and hearing from them about what they love and don't love will make the product better in those first few years. There's a lot of opportunity in the future to open it up and give customers the ability to change what the robot does and deploy their own robots, and we will need to enable that to unlock massive scale. But in those first few years, we'll definitely be hands-on because we want to make sure customers get the value they expected.

[00:33:09] Brian Heater: Something I wanted to loop back around - something we were actually talking about before the camera started - was reinforcement learning. This certainly relates to the conversation around Spot and the App Store, how much has to be pre-programmed into the system. A lot of these conversations around physical AI right now are about what needs to be hard-coded versus what can be done in AI. Looking at the system going forward, how many of these initial applications do you think ultimately are going to have to be hard-coded onto Atlas?

[00:33:47] Aya Durbin: None of them. Part of when we talk about building fundamental tools that make it easy to deploy the robot and easy to train the robot - those are all AI-based tools. Just because we're building foundational technology doesn't mean we're building a hard-coded version of sequencing. We are building all of our applications on an AI-based system. But we're making sure that AI-based system can be directly instructed by an external customer system.

If a customer says, 'I want you to pick this variant that's in this location on the floor and put it into this slot on a dolly,' or they say, 'I want you to find this part in a warehouse and go put it onto this jig at this machine in this other area of my building,' we can do that. We can execute exactly the way the customer wants. It's layering those two things together that is going to be crucial to having a good deployment. But it's all an AI-based system that we're building.

[00:34:43] Brian Heater: Something that's just really incredible, talking to you, talking to people at this company, talking to people in physical AI generally sort of at the forefront of all this, is how much of it hasn't been figured out. It's kind of wild to be at this stage where we're talking about early deployments of the system and some of the really deep, important things still haven't been cracked.

[00:35:14] Aya Durbin: Yes. It definitely does keep me up at night sometimes. There are a lot of problems that haven't been solved, and being at Boston Dynamics, I'm pulled more and more into the dreamer category every day, but because our dreaming here is really based in reality.

There are parts of the problem that haven't been solved yet. We are heavily focused on the research front on making sure we're focused first on the core research problems that impact those early customer deployments. There are an unlimited number of research problems you could solve with humanoids. Customers want this robot to do everything, and my job is to focus us first on the core research problems that actually block us from providing value to customers and ignore the ones for now that we could do. We could make our robot climb up a pole, we could make it climb up a ladder, but is that the most important research problem to be focused on right now? Probably not when it comes to those early customer deployments.

The way I handle that as a pragmatist is focusing us first on the core research problems that you can't succeed unless you solve first, and then kind of layering in the longer-term ones over time.

[00:36:35] Brian Heater: I'm trying to envision what a flowchart would look like as far as industrial customer problems, all of this research. Obviously you and everybody else is keeping up with the latest research. Things are changing overnight and causing everyone to rethink how we're solving these specific problems. You're juggling a lot of plates at once. How are you making sure that all of these disparate things are functioning as part of the same funnel?

[00:37:14] Aya Durbin: We have an amazing leadership team. Alberto Rodriguez, who I've done interviews with before, leads a large portion of our behavior research team on Atlas, and that really helps keep the team focused on our long-term goals. All of our research threads relate to our vision for what this robot will do in the next three to five years. While we have multiple research threads going, they're all tied to a core 'why,' a core problem we're trying to solve, or a core customer we're trying to service.

Having a really strong head of behavior really helps. Having really strong leaders that lead each of those research projects really helps, and the people that work on those teams love keeping up with the latest research. They are some of the best researchers in reinforcement learning, in behavior cloning, in dexterous manipulation in the world, and this is what they love to do. Having them all be the best in their field and be under someone who's so passionate about solving real problems for customers helps a lot, especially as a product person.

Having those collaborators on the team - whether it's Alberto, Chris Bentzel, we have some really great engineering leads on the team - really helps keep the team focused on getting this robot out into the world, because ultimately that's what all the engineering team wants to see too: all their hard work being used by people.

[00:38:44] Brian Heater: As we were walking around, Nick was talking about scaling and production and assembly, and I think 30,000 was the number tossed around as far as how many are going to be produced. When you're specifically thinking about three to five years, what are you thinking about?

[00:39:04] Aya Durbin: Our plan for deployment of Atlas is to start deploying actual pilots out in the world with customers in 2028. That's when our robots, we say, will be members of your team and doing the hard work that we've been talking about. When it comes to the scale numbers, Hyundai has committed to building 30,000 Atlas robots a year starting in 2030, and that's really a mass production version of the robot that is designed to be really easy to build and gets to cost and really gets to customer ROI.

We'll start piloting earlier than that, and we are certainly already on the journey now of getting our robots out into the real world and getting feedback from Hyundai and other customers that we're working with.

[00:39:47] Brian Heater: Do you think you have to give yourself some extra buffer there? When you take a flight, they always make the landing time a little bit later just to give yourself some extra time so they can tell you that you landed a little bit earlier. There are so many moving parts, so many variables here. It's a fool's errand to try to think about three years from now in terms of AI and robotics. There's got to be at least some acceptance that things are going to shift, things are going to move around a bit.

[00:40:23] Aya Durbin: They could, but they could also move closer. I feel more confident than ever that the 2028 timeline is more than feasible, because we're choosing to do a set of tasks that aren't the easiest tasks to do in industry. We're not choosing to do the simplest thing first. We're choosing to do something that we think requires us to build the fundamental tools that enable us to do many things in industry.

When we talk about 2028, we're not talking about doing the most simple thing in a building in 2028. We're talking about doing something that's complex, that's hard work, that's repetitive, that's hard on the body, that requires real exception handling and being truly integrated into your customer system. It's not that we couldn't do work - we can do all types of work now. We don't need to wait till 2028 to pilot. We could pilot right now, but we want to work on meaningful work when we go out into the world with customers. We want to work on tasks that are challenging enough that when we launch, we're launching, and we can then go do lots of other similar types of tasks in the world.

I'm not super worried about the timeline because we're already doing this type of work today. All we're doing right now is improving product performance, improving reliability, improving capability - adding features and functionality for humans to interact with the robot in a more consistent and user-friendly way, adding in tools to deal with exceptions, and hearing from customers about what's not working about those tools. We're not shooting for something that's impossible by 2028. If you have me as a pragmatist saying I feel comfortable with 2028, I feel good about 2028.

[00:42:14] Brian Heater: Of course. Hardware is hard.

[00:42:16] Aya Durbin: It is. The data set's different, and I think everyone has a different expectation. When you say launching, when you say piloting, when you say we're going to have these robots in the world, everyone has a different expectation of what a proof of concept is, of what a pilot is, of what a launch means. For us, it means starting to add real value in a customer facility. For some people that means millions of robots in the world by 2028 - to some people it means two robots.

I feel very comfortable with the timeline of 2028. What I'm more concerned about is the public's perception of these timelines - do they think that, like an LLM, robots are just going to be wandering around the world in 2028 doing anything and everything? That really isn't our first target. First target is do the hard work in industrial environments first.

[00:43:13] Brian Heater: That ship might have sailed. That messaging is going to be real hard to get the cat back in the bag on. But when you say meaningful, what do you mean by meaningful deployments?

[00:43:22] Aya Durbin: Meaningful deployments meaning the robot is doing that hard work that has high turnover, that's hard to hire for, and it's doing it well. Specifically things they would want to get a robot for. Like: I couldn't automate this, and my whole team wants this task automated, and we're bringing a humanoid in to automate the task, and the robot is doing it reliably and passing site acceptance testing and getting through the throughput it needs to do, and the team is looking around saying, 'I like Atlas as a part of my team.' That's the goal for the product.

[00:43:52] Brian Heater: As a product person, you're thinking a lot about ROI. Where does that enter the conversation, and how quickly will people expect these robots to really be proven to be a valuable commodity in the workplace?

[00:44:22] Aya Durbin: I think it's immediate. If we say we're going to launch in 2028, you want to be providing positive ROI to the customer in 2028. That's my goal, and that is the goal of Boston Dynamics. I don't think customers should be deploying robots without an expectation of getting value from them, and I don't think that the industry will survive if customers aren't getting positive value from the system.

Obviously if you deploy a robot in 2028, it's going to take more than a few weeks to get positive ROI, but you should be starting on that journey once you have that successful path of site acceptance testing and proving that your product provides positive return on investment.

[00:44:59] Brian Heater: The thing I've really realized over the past few years is that the humanoid companies that are really going to withstand the test of time are the ones that are going to be able to wait it out until that period where they can really prove out those things.

[00:45:20] Aya Durbin: I totally agree. It's not a sexy conversation, but it's part of building any scalable robotics business. If you've looked at the companies that have had success - even looking at Spot or Stretch - we have to prove that our products at Boston Dynamics provide ROI every day.

The customers we talk to about Atlas - many of them are Spot customers or Stretch customers already - and they expect ROI in a timeframe, and that's not unreasonable to us. It's just not widely discussed right now because we are early in the process of building the technology, and there are a lot of fish to fry on the research front before we get to high amounts of scale for humanoids.

[00:46:06] Brian Heater: Well, I think we're out of time. Aya, thank you so much.

[00:46:09] Aya Durbin: Nice to meet you. Thank you so much.

Brian Heater: Thanks to Aya, Nick, and the rest of Boston Dynamics for accommodating us. Great conversation and excellent tour of your facilities. Aya will also be appearing at our Humanoid Robot Forum event in Chicago. That's happening June 23rd and 24th. Thanks to you as ever for watching. If you've been enjoying the show, please like and subscribe to it and the newsletter of the same name over at automated.fm. And with that, we will see you next week for another episode of Automated.