Interview Transcript
Q1 Could you tell us your name, and when did you first begin working on humanoid robots?
Sangbae Kim: My name is Kim. I’m a professor at MIT. I started working on humanoids in my lab in 2019.
Q2 Did you work on humanoids before that?
Sangbae Kim: Not particularly humanoids. I built quadrupeds ever since 2008. I worked on a gecko climbing robot in 2006. Before that I was working on a robot in 2002.
Q3 What got you into humanoids?
Sangbae Kim: I was just curious about bipedal locomotion because I’d kind of done enough in quadrupedal locomotion. So it was mostly about locomotion.
Q4 Do you have some key publication you’d like to highlight from your early days?
Sangbae Kim: Oh yeah. We had an MPC-based controller and also a reinforcement learning-based controller. So our humanoid is mostly just a platform for locomotion study.
Q5 Do you have a paper title you’d like to mention?
Sangbae Kim: I don’t remember. We can pull it later.
Q6 What were the major challenges you faced in your early days working with humanoids?
Sangbae Kim: I think, in order to minimize the rotational inertia of the leg, we put a lot of the actuators, motors, and gearboxes up near the body, and we needed belts and other transmission mechanisms. System identification was a little bit hard. There were a lot of mistakes, and achieving good accuracy was difficult.
Q7 How did your work evolve over time in terms of design, themes, or research questions?
Sangbae Kim: We didn’t actually change the design that much. Controller-wise, we initially designed an MPC controller—a linear model predictive control version—to make it work. At the same time, we also worked on reinforcement learning. We immediately saw more problems with reinforcement learning than with MPC. MPC is composable. We can do many different things with it. Reinforcement learning is very good at doing one thing only. It’s very difficult to naturally grow the capability from one task to two tasks and beyond. There’s no good way to combine them. If you add more tasks, the overall performance drops. With MPC, you can actually grow capabilities very well. But once you step outside the assumptions of the model, it just can’t perform well.
We’re still working on a framework where we can combine MPC controllers inside neural networks so that reinforcement learning can use them. We’ve had some success with training. Now we’re moving to the next step where we can actually backpropagate through the MPC controllers. We can even modify many parameters, including the gait schedule and so on. We’re almost there. That’s really how our work has evolved: combining MPC and reinforcement learning.
Q8 Has the current emergence of LLMs affected your work at all?
Sangbae Kim: Not at all, because we’re focused on locomotion. I don’t think language models can really help with locomotion. Those are different challenges.
Q9 How do you see the relationship between a robot’s body and its controller?
Sangbae Kim: Sorry, I don’t quite understand what you mean.
Q10 The body and the controller?
Sangbae Kim: The controller always controls the body. Is that what you’re asking?
Q11 I think some people talk about co-evolution between morphology and control.
Sangbae Kim: We didn’t change the design at all. We’ve been refining this design for over 20 years. We invented a proper actuation system that’s actually used throughout the Chinese robotics community now. Even U.S. companies are copying ours.
We know how to build a robot that works, so we didn’t change the design. But the controller has evolved in many different ways. We still don’t have a full answer about which approach is best.
For locomotion, there’s a good chance that some form of reinforcement learning will win, but none of us really know yet. I think MPC can still play a big role, but it might not be enough by itself.
Q12 Can you mention some influential researchers or labs we should check out?
Sangbae Kim: For MPC and reinforcement learning, we read a lot of papers from ETH. For MPC more broadly, we look at the Rugged papers and also Mike Tesch’s papers.
For MPC, much of it is our own work. For reinforcement learning, we were influenced by Marco Hutter’s group, and now everybody is doing similar things. That’s pretty much it. If you go further back, we also read a lot of the biomechanics literature on human walking. Those papers were really helpful. I actually have a very strong background in biomechanics. I read almost every biomechanics paper on human locomotion, so I know that side very well too.
Q13 Do you have any particular favorite paper?
Sangbae Kim: I don’t remember the titles. There were lots of papers. I spent probably half of my reading time on biomechanics literature, but that was back between 2002 and 2010—more than 15 years ago.
Q14 Have you had any unexpected successes or unexpected failures in your research?
Sangbae Kim: There have been so many. We once had a leg explode because the motor was too strong, and the entire robot shattered. That was definitely an unexpected failure. But it actually impressed our DARPA program managers, and we got funding because of it. So it was an unexpected failure and an unexpected success at the same time.
Q15 That’s interesting. To wrap this up, I’d like to ask one last question. What are you most excited about at this Humanoids conference?
Sangbae Kim: I’m very excited that the community is growing. There’s a lot of development happening with sensors and robotic hands, and I’m seeing many new emerging solutions. The progress isn’t incredibly fast, but there are so many people working on it now that I can look around and see development accelerating. That’s very exciting. I think in the next couple of years we’re going to arrive at solutions that people can actually use. MIT basically gave everyone a solution that people still use today in terms of the general form factor. That was an exciting moment. I don’t think humanoids are quite there yet, but I can see that they’re going to come together over the next few years. I’m talking about a very general version—not saying they’ll do everything humans do—but I think the field is starting to converge. If you look at the early development of cars, there were designs all over the place. Eventually everyone evolved toward certain common patterns and paradigms. Even today there’s still a lot of variety in cars, but there are commonalities: how many wheels they have, where the doors are, how they open, and so on. I think we’re starting to see that same kind of evolution in humanoid robotics, and that’s very exciting. The amount of work being exhibited at this conference is really impressive.
Interviewer: Yeah. Thank you so much.
Sangbae Kim: No problem.