Interview Transcript
Q1 Could you tell us when you first began working on humanoid robots?
Jonathan Hurst: I think I first started working on humanoid robots in 1997, maybe when I started school. It’s something I’ve always been interested in understanding. I wanted to know how to build robots that can move the way that people and animals do.
Q2 What was the major challenge in the days when you started working on humanoids?
Jonathan Hurst: Even then, at least for me, I came from a hardware background. So it was very quickly clear that the actuators were insufficient and really needed to change. At that time, the challenge was the transmissions and actuators—making a machine where it’s even physically possible to do the movements that we want. Even if we could figure out how to control it, we first had to have the hardware do it well. I think there were many challenges, but that was the angle that I took from the beginning.
Q3 Would you say the actuators are still a major challenge?
Jonathan Hurst: The actuators are certainly still a major challenge, but it’s something that I feel we understand now. It’s gone from being a basic science question to an engineering problem. We know what we need, and now we keep iterating to make them better and better.
The problem has shifted. We’ve now built machines that can do walking and running behaviors quite well. Now we have to control them well, and that’s what has led us toward reinforcement learning approaches.
We’ve been through different eras of controls. We started by writing heuristic equations for simple models to control a robot to hop around. Then we moved to whole-model-based control with inverse dynamics and all kinds of complexity. After that came online model-based control.
Now, with reinforcement learning and data-driven approaches, it’s just a policy running on the robot. It’s much more computationally efficient, and it can capture highly nonlinear behaviors and exploit the full dynamics of the hardware. Things are starting to come together.
Q4 How much have LLMs in general affected your work?
Jonathan Hurst: I really thought it would take decades to get enough physical context and have these robots start moving and doing useful things before they could really understand anything about the world around them.
It turns out that having an entire internet full of language and images means it’s kind of coming from the top down. Instead of taking decades, we get to meet in the middle.
We still have to figure out the hardware. We understand building from the bottom up, and we still have to engineer a good hardware system. We still need good dynamic control—the physical AI part of it. But then it’s really about the interface. What kind of information needs to go up to more cognitive intelligence, and what context about the world needs to be passed down to the physical AI so it can really operate effectively in the world?
So it actually seems easier than it did.
Q5 Is there any publication that you would like to highlight from your early days?
Jonathan Hurst: Oh boy. There’s an article that I wrote for IEEE magazine where I really laid out some of the challenges of actuation and how to design leg configurations. That would be a good one.
Q6 Do you remember the title?
Jonathan Hurst: No, I’d have to look it up.
Q7 How would you say your work evolved over the years?
Jonathan Hurst: At the very beginning, it was very much hardware. It was about designing physical devices that could have the right leg configurations and the right actuators.
Over time, it evolved toward controls and then AI approaches to control. It’s really interesting how our background in the physics of mobility and locomotion has helped shape how we approach AI as well. Having the right physical intuition lets us write meaningful cost functions and recognize whether a learned solution is actually on the right track.
As much as we’d like these systems to be fully automated, they’re still very much human-in-the-loop. They’re powerful tools that help engineers build better systems.
I want to add something to that. Technically, my work went from hardware up through software. But now, after starting a company, it has shifted toward a much bigger picture. Many of the hardest problems are no longer purely technical.
We have to understand how to build and manage a large organization. We have to determine how we’re going to support robots in the field. We have to develop a safety case and work with insurance companies to understand the risks associated with our robots. We have investor conversations.
It’s a completely different mindset from academia. Running a business is very different from solving an interesting research problem. That’s been a major shift and a significant learning curve.
Q8 I imagine engineering problems are so different from working with people in the real world. How do you educate a CEO or people who can take insights from the real world, from customers, and put them back into your design or engineering?
Jonathan Hurst: We actually have an independent product organization. We have several people whose job is simply to understand the customer applications, understand our technology, find the right match, and write down the list of requirements.
There are hundreds of specific requirements about the robot—its mass, what kind of dust rejection it needs, temperature limitations on surfaces, but also behaviors. For example, it needs to be able to step off an unexpected bump on the ground and tolerate that. We also define all the different ways in which a robot can fall, what amount of damage or failure is acceptable when it does fall, and how durable it needs to be.
All of those requirements for how the robot will be used get translated into engineering requirements. That’s really important because the engineering team is large enough that the people working on the arm may not have the full context of what the torso team is doing. They all understand the interfaces, and they all understand the requirements that apply to them.
That kind of process is necessary when you have a large engineering organization building a complex system into one unified product.
That’s very different from a research lab, where most people can keep the whole system in their head and know what everyone else is doing. It’s much easier to coordinate and brainstorm in that environment. As the project gets bigger, it becomes a different kind of challenge.
Q9 Based on your experience implementing work in the real world, what kind of feedback would you like to give back to academia about the way robotics research is currently done?
Jonathan Hurst: Honestly, the field has already been shifting more and more toward doing things physically on hardware, which is really important. In the past, many people were doing work only in simulation and reporting simulation results, which I tend to discount. Building things, testing them physically, and demonstrating them in reality is already becoming much more common in academia, and that’s a good direction.
Let me add one more thing.
I think a bad habit in academia is pursuing a problem simply because it’s interesting. I think researchers can have a much bigger impact if they’re solving a problem that people actually have—if there’s a clear connection between the research and the real-world impact it could create.
That’s difficult because many professors have gone straight from undergraduate studies to graduate school, then to a postdoc, and then into academia. They may not have firsthand experience with customer problems.
I ran into this myself when trying to understand logistics warehouses. How do they actually work? What do they really look like? You don’t know until you’ve spent time there and maybe even done the job yourself.
I think it’s really important for graduate students, professors, and researchers to understand the domains they’re working in and how their work will apply in the real world. Making that connection to relevance allows you to solve the same interesting problems while making them directly applicable.
Q10 So maybe engineers could do field work?
Jonathan Hurst: Or simply spend time with the people who do those jobs. Ask them questions. Go to a trade show instead of just a conference, and talk with the people there who don’t care whether it’s a robot or not—they care whether their bins get moved or their boxes get sorted.
You’ll develop a much more realistic understanding of the problems. Then you can target your research much more effectively and solve the same interesting technical problems, but in ways that are relevant to real applications.
Q11 You touched on this a bit, but could you give us a rough timeline of your work? You started in 1997. Was there a shift around 2010 and 2020?
Jonathan Hurst: 1997 is when I started college. I went to Carnegie Mellon because I wanted to join the robotics club and compete in the walking machine competition.
During my first year, one of my projects for General Robotics—Howie Choset had just started the robotics minor—was to build a small biped robot using servos and machined aluminum. It didn’t work very well. It could stand and sort of shuffle, but it was nothing like what I imagined a truly athletic machine could do.
That was the beginning of realizing that the problem was much harder than I had imagined, but also much deeper and more interesting. It’s been an entire career spent exploring that problem, and we’re still learning.
As a freshman, I built that first simple walking machine and competed in the walking machine competition. Then I stayed for graduate school and built Maple and Thumper. Maple was a bipedal robot, and Thumper was a monopod. We spent a lot of time studying high-speed cable drives as very low-friction, high-efficiency transmissions.
Cable drives are excellent in some ways, but they also have drawbacks. That’s exactly what research is for—trying ideas that haven’t been explored before. We essentially did a depth-first search on cable drives, explored all of the pros and cons, and documented what we learned. That became a major part of my PhD thesis.
As a result, I now know when cable drives are the right solution and when they aren’t.
Then I came to Oregon State as a professor and continued pursuing the same question: can we build a machine that actually moves like a person?
That ultimately led to ATRIAS, our bipedal robot, which was able to walk and run outdoors before almost anyone else could do it. It was the first robot to really capture walking gait dynamics. Marc Raibert had already demonstrated running robots, but beyond that there really weren’t many robots operating in this space.
ATRIAS could smoothly transition between walking and running while remaining bipedal. That was the point where we decided to start the company. We felt that we understood enough of the core science to be confident that we could build robots capable of going where people go.
From there, the question became: what becomes possible once robots can go where people go? How can we provide value to customers, create a business people are willing to pay for, and build the flywheel that lets us keep improving the technology?
Q12 What’s your vision for the next five or ten years?
Jonathan Hurst: Over the next five to ten years—and really beginning late next year—we’ll be deploying Digit Version 5. It will be the first dynamically stable robot with all of the functional safety features required to operate outside of a work cell. No other dynamically stable robot can do that today.
That’s the scaling moment. Instead of selling just a handful of robots that have to remain inside work cells, we’ll be able to deploy hundreds and eventually thousands of robots that can simply perform their jobs without requiring dedicated infrastructure.
Our focus over the next five years is to build that flywheel: robots operating in the field, collecting data on successes and failures, bringing that data back, and continuously improving the system.
From there, we’ll expand toward better dexterity, greater safety, operating safely alongside people, and manipulating a much wider variety of objects. We’ll start with moving boxes and each-picking, then expand into much more general tasks beyond that initial beachhead market.
Q13 Speaking of software and hardware, how do you see the relationship between the robot’s body and the controller?
Jonathan Hurst: The relationship between the body and the controller is holistic. They have to be designed and integrated together.
Even something as simple as designing the foot of our robot illustrates this. Right now, the foot is nothing special—it’s just a piece of metal with a rubber sole—but it needs to become much better than that.
When we design a foot, it has to provide the right physics for walking while also being easy to simulate so that our reinforcement learning approaches in simulation closely match reality.
That was actually the philosophy behind ATRIAS. ATRIAS stands for Assume The Robot Is A Sphere. The whole point was to make the real robot behave so closely to the simple mathematical model that we didn’t have to worry about all the complicated inertia, friction, and other effects. The controller operated on a reduced-order model, and it worked directly on the physical robot.
With Digit, it’s the same philosophy, except now we use a very high-fidelity simulation of the robot interacting with the world. We perform reinforcement learning in simulation, and transferring those policies from simulation to reality works much better if there’s a close correlation between the two.
So it isn’t just about improving the simulation. It’s also about improving the hardware so that it’s easier to simulate accurately.
Q14 Have you had any unexpected success or failure in your research or in real-world applications?
Jonathan Hurst: I’ve had uncountable unexpected failures. Not nearly as many unexpected successes.
Usually, success is something you dream about, plan for, and work extremely hard to achieve. Successes rarely happen by accident.
Failures, on the other hand, happen constantly. It’s like paddling upstream. Then, when you finally reach your destination, it feels great because you had to overcome so many challenges along the way.
Q15 Do you have any funny episodes you recall?
Jonathan Hurst: Probably, although I hadn’t really thought about that beforehand.
One example is when Digit was about to walk out on stage during a presentation. Our CEO was speaking and was just about to introduce Digit when, backstage, the robot failed to start and fell over.
The team immediately started working frantically to figure out how to reboot it and determine what had gone wrong. The controls team knew the robot was supposed to walk out any minute, but they didn’t yet understand why it had failed.
Meanwhile, the speaker on stage was confidently saying, “The robot walks out. Of course it works. This is easy, right? No big deal.”
That was pretty representative of many of our demonstrations in the early days of the company. Fortunately, the robots are much more reliable now.
Q16 Compared to what you and the robotics community were trying to tackle back in the 1990s, what are the key differences in the challenges today?
Jonathan Hurst: They’re completely different.
In the early humanoid community, most people were building Zero Moment Point (ZMP) humanoids. My own work focused on dynamics, physics, and bipedal locomotion, but the broader community was building robots that were essentially stylized walkers. They relied on position control, flat and predictable ground, and moved very slowly.
I think many people recognized that this wasn’t really human-like movement and probably wasn’t the right long-term approach, but nobody yet knew what the alternative should be.
People started with those systems and then worked on other important problems, such as perception and higher-level system integration—for example, figuring out how to move a chair out of the way before walking through a space.
Researchers were using the tools they had available to solve the problems they believed humanoids would eventually need to address.
The broader progress of robotics has naturally been reflected in the evolution of humanoid robotics.
Q17 Throughout the years we’ve had this intuitive idea of what a humanoid robot is. What would you say makes a robot a humanoid?
Jonathan Hurst: I usually avoid trying to define it.
In fact, I once gave an entire talk about what walking and running actually are. One interesting realization from the dynamic walking community, working alongside biomechanics researchers, roboticists, and control theorists, is that walking and running are really part of the same behavior. One model can reproduce both. So what exactly is that behavior? It’s certainly not the dictionary definition. We don’t even want our robots to “walk” according to the dictionary definition of simply putting one foot in front of the other while moving forward. That’s not sufficient.
I actually coined a term for it. I described it “as compliant interaction with the ground”. It’s a self-stable behavior, somewhat analogous to pendulum dynamics. It’s a special region of dynamics where the swing phase is dominated by inertial forces, the whole body behaves inertially, and the contact dynamics become especially important.
There are a number of other characteristics as well, but it’s honestly difficult to describe. The talk didn’t land particularly well. The audience’s reaction was basically, “Well, I guess he’s got tenure, so he can say whatever he wants.” People didn’t really know what to make of it. They said, “Fine, but how do we actually achieve that?” At the time, we didn’t know. There wasn’t a clear path. So if I can’t even clearly define walking and running as behaviors, it’s even harder to define a humanoid. Roughly speaking, I suppose it means something that’s human-shaped. But for us, a humanoid is a result rather than a goal.
Q18 Do you think there’s a huge gap between the public perception of humanoids and what people think in academia?
Jonathan Hurst: Sure. I think many people imagine a humanoid as something that’s as close to a human as possible—a lifelike machine with a face that’s an almost perfect copy of a person. That’s not what we’re trying to build. That doesn’t mean it’s not an interesting or worthwhile research direction. It’s just that “humanoid” is a very ill-defined English word.
Q19 Could you point to some labs, researchers, or developers that people interested in humanoid robotics should know?
Jonathan Hurst: I’m a little out of touch with many of the newer labs that are doing excellent work. At this conference, I’m seeing talks by people I haven’t necessarily met before, and they have very good things to say. I don’t think I’m the best person to highlight the new up-and-coming researchers.
Q20 Who or what motivated you to work on humanoids in the first place?
Jonathan Hurst: For me, it’s about movement and motion. Think about two things that many people are fascinated by. On one hand, there’s human movement. People enjoy choreography, dance, kung fu movies, sports, athletics, and beautiful movement. From an evolutionary perspective, we’re incredibly good at detecting when movement is just slightly off. One theory is that if you’re looking at a herd of animals, you can identify the one with the slightest limp, and that’s the one you pursue. We’re naturally good at recognizing movement—when it’s right and when it’s wrong. People love it. It’s part of art. On the other hand, many engineers love machines: steam engines, mechanical clocks and watches, engines, cars, suspensions—interesting mechanisms that work beautifully. Now combine those two things. That’s robotics.
We build machines that move in beautiful ways. They move in ways that look natural and feel right. There’s an aesthetic to natural movement. That’s what has always motivated me. We’re still pursuing it. We caught glimpses of it when ATRIAS first started running. To me, that looked like natural movement. I thought it was beautiful. We saw it again with Cassie, using reinforcement learning to produce running gaits. It looks like an animal. I’m not sure what animal, but it looks natural and beautiful. We’re now working toward having our humanoid robots move that way as well—even when they’re carrying bins or moving boxes. You can imagine a robot reaching over, grabbing a bin, using a little bit of ballistic motion, accelerating smoothly, sliding across, and placing or even tossing it naturally. Once these machines move in ways that are robust, efficient, and fast, I think they’ll naturally begin to look right, because those qualities lead to the correct solution. The same idea applies to making a humanoid robot look like a person. I don’t want to copy or puppeteer human movement or prescribe trajectories that I believe look natural. Instead, I want to optimize for the hard metrics—robustness, speed, efficiency, and everything else that really matters. I believe that if we solve those problems well, natural-looking movement will emerge as the result.
Q21 Is there anything else you’d like to highlight from your work?
Jonathan Hurst: Oh, I don’t know. That’s a pretty good list of questions.
Q22 If you think of anything later, please feel free to send us materials or anything. To wrap this up, I’d like to ask one last question. What are you most excited about at this Humanoids conference?
Jonathan Hurst: Every conference is different. Right now, this is a very fast-moving field. Every year we’re all learning an enormous amount, and we come back together to share what we’ve discovered. Each conference usually develops a theme—something the community has collectively learned, figured out, or solved. I don’t know what this year’s theme is yet. I’m excited to hear the talks, have the conversations, and discover what we’ve collectively reached as the next conclusion. Thank you so much.
Jonathan Hurst: Great.