Interview Transcript
Q1 Could you tell us again your name and your background, like your upbringing?
Sethu Vijayakumar: Hi. My name is Sethu Vijayakumar. I’m a professor of robotics at the University of Edinburgh and one of the founding directors of the Edinburgh Centre for Robotics. I also have a second role as the program director for AI and robotics at the Alan Turing Institute, which is the UK’s national institute for AI and data science. I have a very interesting background because I originally come from India. I did my undergraduate studies in India, and then I did my master’s and PhD in Japan at the Tokyo Institute of Technology. After that, I did a year and a half of postdoc at the Brain Science Institute. My initial training was in machine learning and mathematics. I did my robotics work during my PhD.
Then at the Brain Science Institute, I was very fortunate to work on human motor control—trying to take insights from human motion, balance, and control and apply them to humanoid robots, and understand how machines and humans can learn from each other. After Japan, I had my first faculty job at the University of Southern California in Los Angeles. I spent about five years at USC working on everything related to applying machine learning to robotics, at a time when robot learning for humanoid robots was not very fashionable yet. We did some of the early work applying what we called local weighted learning techniques—locally weighted regression techniques—for things like system identification, inverse dynamics control, and getting whole-body systems working for the first time in locomotion and manipulation.
After USC, I’ve been at the University of Edinburgh for almost 20 years, working on various things related to robotics, machine learning, and the new trends in applying AI—not just for locomotion but across the whole robotics stack. Edinburgh recently celebrated 60 years of AI, so it is one of the places steeped in AI history, which is very exciting.
Q2You went to graduate school in the 1990s?
Sethu Vijayakumar: I did my PhD. I finished my PhD in 1998.
Q3 What was it like to do a robotics PhD in Japan?
Sethu Vijayakumar: It was very different from now. At that time, robotics was still a very exotic field where people thought of robots as something that only existed in very exclusive labs. People had access to robots only in very well-funded laboratories. The computing power required to make some of these robots work almost filled an entire building. Some of the hydraulics I initially worked with—some of the Sarcos robots, for example—were among the first and finest hydraulic robots. It took us forever just to get basic sensing, control loops, and actuation working because everything had to be routed. There was very little embedded systems capability, so everything had to go through very specialized compute cards. We had some computers doing simulations at that time, but in a way every robot project was a whole mini-project in itself, going from drivers to hardware to software. I think PhD students today have it much easier because you can download things from GitHub. You have standardized drivers, and you can get a robot arm or robot leg working in half a day. It took us a year to get something like that going.
Q4 How did that training and hard work shape you as a researcher today?
Sethu Vijayakumar: I would say I understand the details—the nitty-gritty—of some of the complexities of robotics hardware and control loops much better than some of even my PhD students because you had to do everything from scratch. You would still do simulations at that time. We used MATLAB, Python, and specialized Sun animation hardware and graphics hardware. We did a lot of simulation work. But the sim-to-real gap was much bigger than it is right now. You would get everything working in simulation, but the moment you put it on hardware, you had huge sensor noise, lots of latency, and we had to write specialized kernels in our operating system to deal with this and ensure low latency in command streams. This was all pre-ROS, pre-embedded systems for sensing, pre-LiDAR. Everything was either a combination of IMUs with vision-based tracking, or it was basically AR-tag-based systems with very good marker-based tracking. So for me, getting that trial by fire—understanding what it takes to get a robot to walk—makes me appreciate the actual intricacies much more.
Q5 What advice would you give to students today getting into robotics with all these new technologies and AI tools?
Sethu Vijayakumar: I would say absolutely exploit all the innovations that are coming out. I would never say don’t use AI tools, ChatGPT, or other tools for coding. Absolutely use them to the fullest because there is no point reinventing the wheel. But at the same time, when you use these things, you need to have a very solid understanding of the physics of what it means to be working in a particular function space. When you’re doing locomotion, when you’re doing manipulation, even if you’re doing end-to-end manipulation, you still need to understand forward and inverse kinematics. You need to understand affordances. You need to understand what it means to reach for something and grab a box. Just because end-to-end methods work reasonably well for most cases, except the edge cases, that does not mean you should ignore how inverse kinematics solves those problems. What I see is that current trends still leave a lot of gaps in terms of good data. What happens if there is no data in part of the space you’ve not explored? Reliability and creating systems that are verifiable and explainable become even harder if you don’t have a good foundation and understanding of the basics—the physics, Newtonian dynamics. Why would you try to learn gravity from data? It’s all about inductive biases. How do you put these physics insights and knowledge about robotics into modern machine learning techniques?
Q6 What was it like when you moved to the United States and Southern California?
Sethu Vijayakumar: I really enjoyed my time in Japan because of all the cultural experiences—the festivals, the seasons. California was very different in many ways. The weather was great. It was consistently sunny. We used to live by the beach, so it was wonderful. The work culture and ethos were also extremely different. I was really exposed to the speed of doing things, which was completely different. We would take big leaps of faith and do blue-sky research much more aggressively. My time in California was all about experimenting and thinking outside the box. Japan was very good foundationally. From a contextual point of view, the foundations are very strong. But I would say research is more risk-averse in Japan compared to the US. Both approaches have their own advantages. When I moved to Europe, I felt it was somewhere in between the two. It was an interesting balance because you could still do things with a very long horizon—thinking about basic problems that would bear fruit in two, three, or four years. But at the same time, there was a dynamism around taking research to market: how do you get VC investment into interesting ideas that come out of our labs? That was extremely exciting.
Q7 Could you talk about your industry collaborations?
Sethu Vijayakumar: Personally, I work with industries all around the world. I’ve worked with many Japanese companies, including Hitachi on internal logistics and moving goods around. We have had a long collaboration with Honda Research Institute in Japan and Germany, and we have also worked with Toyota. Within the UK, we have worked with companies like Dyson, Ocado, and Touch Sensing labs. Most of these labs and research centers were relatively young compared to companies like Honda and Toyota, but they were disrupting markets in areas where the research domain didn’t previously exist. Ocado, for example, was one of the pioneers of grocery delivery using robotic fulfillment in the UK. Dyson was innovating in household technologies and looking at how automation could be taken to the next level. Similarly, Shadow Robot was one of the pioneers of innovation in robotic hands before the current explosion of humanoid robotics. These were niches of innovative industries that created markets where they didn’t exist.
Q8You’ve worked on three different continents with different work cultures and funding structures. How have your research questions changed, if at all?
Sethu Vijayakumar: Regarding funding structures, it’s very interesting. When I moved to Europe, I thought the ecosystem was very unique. The European Union framework programs, the Horizon programs, and now newer versions of those programs are very unique. The individual fellowships and joint Horizon funds are extremely valuable because I made friends and started collaborating with people I would not have worked with otherwise. The European Horizon programs and framework programs were instrumental in bringing together excellence from different pockets across Europe. I would congratulate Europe for setting that up. Compared to the very program-driven approaches in the US and UK, such as DARPA or ARIA, those are focused on grand challenges. That’s also very important. But they do not necessarily force you to collaborate with partners whom you would not have interacted with before. That, in my opinion, is the biggest difference between the US, European, and UK ecosystems.
Q9 Would you say US universities tend to collaborate within departments more?
Sethu Vijayakumar: Absolutely. I think there is a lot more collaboration within departments. I don’t have very long experience working in the US, but based on my limited experience, I felt it was much more directed toward within-department collaboration and interdisciplinary collaborations across departments, but less so across multiple universities, multiple departments, and companies. And I think that is related to how NSF or similar funding is often allocated to one university.
Q10 What do you think is your biggest contribution to this field?
Sethu Vijayakumar: Along with some of my mentors—Stefan Schaal, now at Google, and Chris Atkeson at CMU—we were, almost 25 years ago, doing some of the first work applying real-time machine learning to the control of anthropomorphic systems. This was before there was a field called robot learning. At that time, robotics was very traditional, and machine learning was making tremendous advances in regression, classification, and learning techniques across the board. But nobody had really put those two things together. I think roboticists were skeptical of machine learning because they questioned whether it could provide the guarantees and safety required. At the same time, machine learning researchers often did not understand the real problems in robotics. People would take an existing algorithm and say, “That’s regression, let’s just plug it in and solve the robotics problem.” But often the important contextual details were missing. I would say my biggest contribution was developing tools for applying machine learning to domains and problems that were very specific to locomotion, manipulation, and whole-body humanoid robotic systems, and really adapting machine learning methods to robotic domains 20 years ago.
Q11 How do you see yourself contributing in the next five, ten, or twenty years?
Sethu Vijayakumar: I personally think my next set of research contributions will be about grounding the current AI research—and what I would almost call the hype around AI for robotics—into something where we exploit the best of both worlds. What I mean by that is we still want insights from methods like model predictive control, inductive biases, gravity compensation, and classical techniques. But then we want to fill the gaps in these classical techniques with modern AI methods. How can we provide good initializations? How can we use massive models or large amounts of data to explore much more efficiently? I think my contribution will come from both sides: grounding this work so that robotics does not go in one extreme direction or the other.
Q12 Do you have any specific application areas in mind?
Sethu Vijayakumar: For robotics, we have looked at many different domains. We have worked in extreme environments. By that I mean nuclear decommissioning at Sellafield in the UK. We have looked at space robotics, working with NASA and the Johnson Space Center on projects involving the Valkyrie robot. We have also worked with underwater robotics. So robots for extreme environments are a major area for us. I would say about 30% of our work is in the healthcare domain. For example, exoskeletons—combining human volition, human residual capabilities, and using that to create assistive systems. This includes exoskeletons and functional electrical stimulation systems that can work collaboratively to achieve tasks. Examples include supporting someone who is infirm, rehabilitation after an injury, or helping address the challenges of aging so people can live longer and more independently in their own homes. We have a whole program around assisted living, and our collaboration with Japan through the Moonshot project is along those lines. We are looking at assistive robotics in healthcare, autonomous driving, and systems that can live inside homes while combining robotics with the remaining capabilities and intentions of humans. The idea is not to take humans out of the loop. The human is still driving the system, and the robot is an assistive mechanism.
Q13 Some people would say, “You don’t want humans to assist robots. The whole point of robots is to automate everything.” What would you say to that?
Sethu Vijayakumar: Just to clarify, my view is not that humans should be assistants to robots. My point is about human-centric design for robotics. What that means is that even though robots may have large parts of their workspace where they can perform tasks autonomously—avoiding obstacles, being safe, not bumping into things, not falling over—we do not want to remove human decision-making from the loop. We want humans to remain in control of the things they want to control. It is about seamlessly moving between the two extremes of teleoperation and full autonomy, and being able to move between those extremes even within a single task.
Q14 Is there anything else you’d like to highlight about your research?
Sethu Vijayakumar: The only other thing I would add is related to my role at the Alan Turing Institute over the past five or six years. It has been a very unique experience because all my life I have been an academic doing foundational research, thinking about how research can be translated through market needs and industry. But the Turing role has given me an opportunity to step back and think about technology itself. Is the technology we are generating and creating really necessary? Is it important for society? Is it developed in a way that is equitable and fair? Will it benefit society as a whole? Those questions are not always obvious in the research we do. The Turing Institute has helped me think about technology from the perspective of how it will change society, the economy, social dynamics, and our daily lives—not just for affluent people, but also for those who really need these technologies for assistance. For me, we need more of that kind of thinking in robotics and machine learning. Some people consider those issues as an afterthought, but I would insist that these questions need to be built in at the time of design and project specification. Fairness, equality, and developing technology for social good should be addressed at the beginning of the project, not afterward.
Q15 How would you educate engineers or bring that broader perspective into robotics?
Sethu Vijayakumar: You are absolutely right that when people are deeply focused on robotics research, it can sometimes be difficult to see the bigger picture. I think it falls on senior people in the community to help set directions: where funding should go, what the priorities should be, and how technologies should be evaluated. The responsibility falls on funding agencies, senior leaders in the community, and also on education. When teaching a robotics course, it should not only be about dynamics, inverse kinematics, or reinforcement learning. It should also include questions like:
“If you do this, what problem are you solving?”
“And if you solve this problem, why is robotics the right technology for it?”
“How do we address grand challenges in manufacturing? How do we address challenges in asset inspection?”
“If we use robotics and AI to solve these problems, how will the economics work? Who benefits? How do we ensure that the benefits are shared?”
Those conversations need to be part of teaching. They cannot be an afterthought.
Q16 Since this is a humanoids conference, I would like to ask a very human-related question: what makes a robot a humanoid?
Sethu Vijayakumar: I think everyone here has a slightly different answer. For me, I am not actually obsessed with the form factor. I have worked with many interesting humanoids, but what makes something humanoid is whether the actions of the humanoid are relatable to human actions. Some people frame it in terms of the data humans generate—whether human motion can easily be retargeted and played back on a robotic system. Personally, I don’t think that is the most important thing. I think what is more important is whether a robot can anticipate and predict the actions of another agent in a context where it is interacting with that robot. Can we find intuitive ways of anticipating actions? Humanoid form factors, if programmed correctly, should allow us to become much better at predicting and understanding robot behavior. Then, when we have real physical human-robot interaction, we can do it in a much safer and more natural way.
Q17 So safety is one of the big reasons to pursue humanoids?
Sethu Vijayakumar: I would say so, yes.
Q18 Speaking from your unique position thinking about technology, politics, and society, what do you think about the social implications of humanoids?
Sethu Vijayakumar: Good question. I haven’t mentioned some of the other things I do. I give some TED talks, and I am also one of the judges on BBC Robot Wars. Every time I am in a public-facing situation, I get asked:
“Are robots going to take our jobs?”
“How will the economy change?”
“What new skills will people need?”
From a political standpoint, I think it is important for governments to invest in technologies involving cutting-edge research. But you have to pick and choose your battles because every country is different in terms of size, resources, and access to markets. Countries need to identify areas where they want to excel and invest strategically. At the same time, governments and policymakers need to think about educating the next generation for the jobs of the future. When computers changed society, even creative professions like art and film were transformed. The technology required to make a film 25 years ago is very different from what is required today. Robotics and AI will become part of every aspect of our lives. Politicians and governments need to think ahead and build the right infrastructure and education systems. Countries that do this well will prosper. But it is difficult because there is a long latency—it can take 10 to 15 years to build the necessary pipeline. If you do not think about it now, you will already be late.
Q19 You have a very unique position in academia and also at the Alan Turing Institute, with a public-facing role. As someone with multiple perspectives, what are you most excited about today in robotics?
Sethu Vijayakumar: From my perspective, I understand both the great leaps that robotics and humanoid robots have made and the current limitations. What really excites me, wearing all these different hats, is that we are now able to communicate and democratize insights from humanoid robotics research to the general public. People have been asking: “What is the ChatGPT moment of humanoid robotics?” I think we are getting to a stage where it is no longer only researchers in labs or robotics manufacturers defining how robots will be used. The general public and people who are not robotics specialists are starting to think of clever applications of robotics that could make a difference economically, environmentally, or socially. For me, this is the democratization of robotics applications across different domains. And that is what I am most excited about.
Interviewer: I have asked all my questions, and that’s a fantastic note to end the interview on.
Sethu Vijayakumar: Perfect.
Interviewer: Thank you so much.
Sethu Vijayakumar: Thank you very much. It was a pleasure.