Interview with Yasuo Kuniyoshi

Q1: Welcome. If we can begin by getting some background on you. Could you share your full name, where you were born, where you come from, and anything about your early life, family, or early education that you would like to share?

Yasuo Kuniyoshi: Well, my name is Yasuo Kuniyoshi. I was born in Japan, in Chiba Prefecture. My father was a professor in English education at a local national university, and my mother was also an English teacher. I still live in Chiba now. I have a house there and a family—my wife and three children. They are already grown, but actually, a picture of my family when they were young appeared in Science Robotics. I can show you.

Q2: Yes, please. That would be great.

Yasuo Kuniyoshi: This picture is from the backyard of my house. Here is my wife when she was young, and these are my three children. The picture was used to explain different aspects of imitation—how differently each child imitates the mother. For example, the youngest child here is holding this tool and swinging it around, looking very carefully at the mother. She is really happy because she is reproducing the same motion and posture as the mother. But this is a mop, so she cannot actually sweep. She is imitating the movement rather than accomplishing the purpose. This child here ran and grabbed the correct tool and started sweeping. But he became so excited about sweeping that he scattered all the leaves around, so he was not really accomplishing the purpose of cleaning the garden. The eldest child took a different set of tools and a completely different posture from the mother, but he achieved the same goal—cleaning up the garden. So this picture became rather famous in the community interested in imitation learning.

Q3: Was this connected to a research article?

Yasuo Kuniyoshi: Yes. Actually, it was a review article with Gordon Cheng, Carina Cramer, Michael Beetz, and myself. It was a review article about robot learning, and it appeared in Science Robotics some years ago.

Q4: That will be great to include in the interview footage. Thank you. Could you talk a little bit about your educational background—where you went to school, mentors who influenced you, and the early stages of your career?

Yasuo Kuniyoshi: I graduated from the University of Tokyo. I completed my bachelor’s, master’s, and PhD there. My thesis advisor was Professor Inoue. He has since retired, but at that time he was one of the leading researchers in intelligent robotics. He also led a large national project on humanoid robots. After I graduated and received my PhD, I joined the Electrotechnical Laboratory, or ETL, which was a national institute focused on electrical engineering and technology. Later, that institute was reorganized into what is now called the Advanced Industrial Science and Technology, or AIST. I worked there as a researcher for ten years. During that period, from 1996 to 1997, I went to MIT and joined Rodney Brooks’ group as a visiting scholar. I was a member of what was called the “Zoo” group—not because of animals, but because it was a group of very unusual and interesting researchers from around the world. That was a very interesting experience. Then, in 2001, I moved to the University of Tokyo as an associate professor, and in 2005 I became a full professor. I have been there since then. In 2012, I became a joint director of RIKEN, another national research institute in Japan. We also have the Brain and Toyota Collaboration Center, which is a center connected with brain science research.

Q5: Is that a kind of translational industry connection?

Yasuo Kuniyoshi: Not exactly translational. It is more of a joint research activity. The center is called the Brain and Toyota Collaboration Center. There are several research units focusing on brain science, and we collaborate with Toyota on basic research topics that are also of strong interest to Toyota. I also became an adjunct director of the AI center at the University of Tokyo in 2016. The formal name is the Next Generation Artificial Intelligence Research Center, which I actually established. So that is the rough history of my career.

Q6: Now that we are up to date, I would like to go back in time and ask how you first became interested in robotics. Were there particular books, shows, or experiences that inspired you?

Yasuo Kuniyoshi: Yes. When I was an applied physics undergraduate, I became strongly interested in artificial intelligence. So I started reading books and papers about AI.

Q7: How did you first encounter AI as a concept?

Yasuo Kuniyoshi: Actually, even when I was a child, I had a very strong interest in computers. At that time, computers meant mainframes—large, bulky machines with tapes running around and flashing lights. I was really fascinated by them, even when I was around five years old. Later, during my undergraduate years, I became somewhat lost. I wondered where I should head in the future. I started losing motivation in pursuing applied physics and began thinking about what I really wanted to do. I looked deeply inside myself and asked what was at the core of my interests. Then I remembered that I really liked computers. Another thing that remained with me was the mystery of humans—particularly how humans understand each other and communicate with each other even though our brains are physically disconnected. So I asked myself: what field can address both humans and computers? The answer was artificial intelligence. AI was exactly what I should pursue. So I started studying computer science and artificial intelligence. I had been studying physics, so I did not know much about AI or computer science at that point. I started reading everything I could find in the university library. I literally started from the beginning. There were journals like Artificial Intelligence, and I started from page one, article number one, of volume one. At first, it was very difficult. I did not understand the technical terms at all. But I struggled and continued making an effort. After a while, I could read faster. Eventually, when I read a paper, I could understand what it was about and whether it was interesting even before reading all the details. It was a very good experience. I learned a lot.

Q8: What years are we talking about when you were doing this reading?

Yasuo Kuniyoshi: It was during my senior year of undergraduate study, around 1985.

Q9: So there were already AI journals and a reasonable archive of material available at that time?

Yasuo Kuniyoshi: Yes, but not too much. It was very different from nowadays. At that time, the entire literature up to then was maybe just one shelf or something like that. It was possible to read everything. After reading the literature, I came to the idea that many researchers were doing things like creating clever computer programs—intelligent computer programs where humans provided some input, the program produced some output, and then researchers interpreted the result by saying the system performed some intelligent task. I felt there was something wrong with this. I thought that what we really needed was interaction with the real world, with active exploration and learning. That idea became the foundation of my research even until now. So I changed my major and entered graduate school in computer science, in the Information Engineering Department, and joined the laboratory of Professor Inoue, where they were doing intelligent robotics.

Q10: What was the first robotics project you worked on?

Yasuo Kuniyoshi: Actually, on the first day when I joined the laboratory as a graduate student, Professor Inoue asked all of the new students to draw a picture illustrating what we wanted to do in our future research. This is the picture I drew with a Macintosh. What it shows is a robot sitting here with its eyes closed, but listening. Then it hears some noise, opens its eyes, and notices that something has been placed on the table. Suppose there is another person in front of the robot doing this. A human places a block on the table, and the robot observes that action. Then the robot tries to reproduce it. During the process, the robot may fail, but it becomes interested and tries again. It grabs the block and attempts to reproduce the same phenomenon. So this represents multimodal perception, active learning, exploration, and interaction with the real world. I think I did a very good job with this drawing. Even now, I think it was a very nice idea.

Q11: This was a result of reading all the AI literature for a year?

Yasuo Kuniyoshi: Yes, exactly. This was a research proposal. After discussing it with my supervisor, we narrowed the topic down to learning human actions visually by watching and reproducing them. Nowadays, this is called imitation learning. That became my PhD research, and I eventually developed the first visual imitation learning system.

Q12: Could you describe that PhD work?

Yasuo Kuniyoshi: Yes. This slide shows that work. It was completed around 1991. It was not a very fancy system by today’s standards. This is my hand, and in front of the camera I perform a block-stacking task—in this case, building a table-like structure. The important point was that the system recognized each of my actions online and in real time. It classified each step and reported it symbolically, saying things like, “You performed this action now.” That result was stored and then transferred to the robot side. The robot observed the scene and reproduced the same task.

Q13: Was there something particularly clever about that system?

Yasuo Kuniyoshi: Yes. The initial placement of the blocks was different from the demonstration. The locations of the blocks were changed. So the system could not simply memorize the movement. It had to understand the meaning of each action. It had to infer that it needed to find a certain type of block somewhere, locate it, and then perform the task. So, in short, the system needed to understand some meaning behind each action in order to reproduce it. This work was highly appreciated by the research community. I received several awards for it, including the IJCAI Outstanding Paper Award, which is one of the top conferences in artificial intelligence. That was the first major research work I did.

Q14: Leading from your PhD work, I wanted to ask about the main research questions that have guided your career. How have those questions changed over time?

Yasuo Kuniyoshi: My first research question was, of course, visual imitation learning. The question was: how do humans actually perceive and understand the actions of others? At that time, this was almost 40 years ago, and the computational power available was extremely limited. Even processing one static image could take several dozen seconds or even a minute. What I tried to do was process moving images—video—and achieve real-time recognition. This was a very difficult problem. I needed to find a clever idea, a fundamental principle. My supervisor told me: “Humans can do this easily. Think about how humans do it.” So I thought about it deeply. Again, I went to the library and read a lot of psychology literature. From that, I developed the ideas of contextually driven active attention control and anticipation-based action segmentation. These were the core ideas I extracted from psychology research about how humans observe other people’s actions.

Q15: So after understanding how humans perceive and recognize actions, how did your research questions develop from there?

Yasuo Kuniyoshi: The next question was how humans acquire these capabilities—how humans develop the ability to perceive, understand, and act. In short, this became a question about cognitive development. I started looking at child development and began working on developmental robotics. So the second major question was: how do babies develop these cognitive capabilities, particularly imitation capabilities? Then, by digging deeper into this issue, I started asking: what is the starting point? Where does development begin? At the beginning, a system has to start with something. We have to decide what initial seeds should be provided. That led me to the idea that embodiment provides a very important foundation. Also, affordances are important. These concepts—especially embodiment—had been discussed in philosophy, linguistics, psychology, and robotics, but the definitions were still rather vague. The question was: can we quantify embodiment? What are the conditions that make embodiment effective? Those became important research questions for me.

Q16: Recently, you have also been working on aesthetics and morality. How did those topics become connected to your earlier research?

Yasuo Kuniyoshi: Recently, I started asking how aesthetics and morality emerge and develop in humans, particularly in babies. The reason is that I wanted to think about something beyond sensorimotor capabilities or traditional intelligence, such as planning and problem solving. I wanted to address something more connected to humanity. This is becoming increasingly important because of the rapid advancement of AI technologies and their impact on society. For AI alignment, we need to understand how to implement ethical functions and moral capabilities into artificial systems. My opinion is that it is very important to understand how morality begins and how humans construct these concepts and capabilities. Currently, one common approach to AI alignment is reinforcement learning with human feedback. But I think this is a rather superficial approach. The training data can never cover every possible situation. There will always be some loopholes. So my opinion is that a system needs a fundamental generative mechanism for important capabilities like morality. This can only be achieved by understanding developmental principles. There is a long story about why aesthetics and morality connect with embodiment, but in short, embodiment provides the connection.

Q17: You also mentioned internal organs as part of embodiment. How do they connect to cognition and morality?

Yasuo Kuniyoshi: Yes. Internal organs are another important part of embodiment. Many people already know that there is a connection between the body and emotion. This has been discussed since ideas like the somatic marker hypothesis. But the concrete modeling and mechanism are still unclear. So one of my latest research questions is: what is the role of internal organs in emotion, value systems, morality, and cognition? This is something I am actively working on now.

Q18: Looking across your career, you have described a clear progression from imitation learning, to development, to embodiment. Have there been challenges—technical, institutional, or social—that significantly affected the direction of your research?

Yasuo Kuniyoshi: Yes. Actually, there was a very difficult period when I was working on developing a fully dynamic, adult-sized humanoid robot. This was around 1996 to 2003. At that time, we were a very small group—not even really a team at first. It was just myself and another researcher, and later a few more people joined. Building a complete humanoid robot required enormous effort. It consumed a huge amount of money and many years of work. For a long time, the system was still incomplete, with many problems, and we had no papers coming out from the project. It was very difficult psychologically. I was anxious that maybe we would end up with nothing—that we would spend years and resources without producing meaningful results.
But somehow, I endured. Later, I had very strong collaborators, including Gordon Cheng, who is now a professor at the Technical University of Munich. After moving to the University of Tokyo, I also had excellent students who were extremely talented and hardworking. With their help, we finally achieved initial results. We integrated the whole system and succeeded in whole-body dynamic actions, which was the first achievement of that kind in the world at that time. Nowadays, you can see humanoid robots on YouTube doing very impressive things—jumping, flipping, and performing dynamic movements. But 20 years ago, this was a very early stage.

Q19: Were you designing the entire robot from scratch during that project?

Yasuo Kuniyoshi: Yes, everything. We started by asking: what is the meaning of having a humanoid body? One important idea was that if the robot performs motions involving contact with the entire body, then the outer shape of the robot becomes important. Also, if the robot performs highly dynamic whole-body motion, then mass, inertia, and power become critical. The robot needs to be close to humans in these dimensions. So we decided to target what we called the “roll-and-rise motion.” The robot lies flat on the floor, swings its legs upward, swings them downward, and then stands up in one movement. This became our first target action.

Q20: What was interesting about that experiment?

Yasuo Kuniyoshi: Actually, the failures were very interesting. The robot was lying on the floor, trying to swing its legs and stand up. Sometimes it failed, but the way it failed began to look very human. That was one of the important effects of having a human-like body. Because the physics governing the body is similar to humans, the robot’s behavior also begins to resemble human behavior. Eventually, after repairing damaged parts and trying again, it succeeded. The robot was not given a simple fixed motion. We designed the body and the control system so that the physical properties of the robot contributed to the behavior.

Q21: Did this require developing many components yourself?

Yasuo Kuniyoshi: Yes. For this type of humanoid robot, commercial components were not sufficient. The design constraints were very strong. The surface had to be smooth, the dimensions and mass distribution had to be close to humans, and the power had to be sufficient for dynamic actions. So we designed many things ourselves. We developed special motors, drive circuits, communication systems, power supplies, and later tactile sensors covering the body. Everything was custom-made.

Q22: During the course of this research and your career more broadly, were there particular tools—technical or conceptual—that became especially important in shaping your work?

Yasuo Kuniyoshi: As a conclusion, I would say not exactly, because my research has not really been tool-driven. Of course, I enjoy using the latest technologies. For example, reinforcement learning has become a very important technology for us, and we use it. We have some research that relies heavily on it. But it did not define my research direction. My research questions usually came first, and then I looked for the appropriate tools and technologies to address them.

Q23: Earlier you mentioned the difficulty of funding a large humanoid robotics project. Could you talk about where funding for your research has come from over the years and whether those relationships influenced your work?

Yasuo Kuniyoshi: In short, most of my funding has come from the government. When I was at ETL in the early days, much of the funding came from large national projects that supported research activities there. It was relatively easy to obtain some portion of that funding. After moving to the University of Tokyo, I began receiving typical government research funding called KAKENHI, which supports basic science. Since then, much of my funding has come from KAKENHI and other national funding programs. I have also collaborated with industry from time to time, which has become an important part of my research activities, but the major source has been government funding. For national funding, usually you simply see a call for applications and submit a proposal. For industry collaboration, it is different. Companies often find us through our publications or research results. They contact us, discuss possibilities, and then develop joint projects.

Q24: Is there anything you would have liked to pursue during your career but have not had the opportunity to do?

Yasuo Kuniyoshi: There are many things, but if I name just one, probably BMI—brain-machine interfaces. I have always been interested in this topic, but I never actually worked on it because I had many other things to do. Also, I did not have a sufficiently original idea in that area. But it remains something I am interested in.

Q25: What would you say are your most important contributions to robotics?

Yasuo Kuniyoshi: In one word: embodiment. Across several decades of research, I conducted a series of experiments and analyses with robots to clarify what embodiment means in concrete, quantifiable terms. I also wanted to demonstrate the effects of embodiment by showing that embodiment structures interactions with the world, and through those structured interactions, behaviors can emerge. I developed my own definition of embodiment: Embodiment provides constraints on interactions with the world, and those constraints structure the interactions. This sounds abstract, but abstraction is important because it allows the concept to be generalized.

Q26: Could you explain an example of how embodiment creates this kind of structure?

Yasuo Kuniyoshi: In the roll-and-rise motion, there is a very important instant during the action. This occurs when both feet touch the ground. It is a critical moment because the robot must have exactly the right posture. If there is an error at that moment, the robot will fail. Through numerical experiments, we clarified that during the entire movement there is only one instant that is extremely important. The rest of the movement is much less sensitive. So embodiment structures the whole phase space of the action. It creates a meaningful structure where one point becomes critical. This also becomes a basis for communication. If another person is learning the same movement, you can tell them: “This moment is important. Be careful here.”
Then they can succeed. So embodiment creates meaningful information for both performing actions and communicating about actions.

Q27: Is there another example where you have seen this emergence through embodiment?

Yasuo Kuniyoshi: Yes. Another example is locomotion. In one simulation, each leg was driven by a chaotic oscillator. The system itself did not know how to coordinate the legs. However, because of the body structure—the embodiment—the coordinated gait pattern emerged. The robot started walking in a particular direction. This principle is not an artificial phenomenon. It is biologically inspired. In vertebrates, there are spinal circuits called central pattern generators, or CPGs. These consist of nonlinear oscillators and have a similar function. We created robotic systems based on this idea and showed that embodied structures can generate coordinated behaviors. For example, with quadruped pneumatic-driven robots, the system does not explicitly know how to control each muscle, but when activated, it immediately begins producing movements such as walking forward, walking backward, or swinging. The body itself provides the structure that allows these capabilities to emerge.

Q28: How did this connect back to your work on human development?

Yasuo Kuniyoshi: Later, we integrated these ideas into models of human development. We built an accurate musculoskeletal model of the human fetus and combined it with nervous system models and spinal circuits. The same principle applies: behavior emerges under the constraints of embodiment, and this creates structures and information that can be learned by the developing brain. So the overall story is embodied development of human cognition. That is the direction we are still pursuing.

Q29: You mentioned the term “neck point” earlier. Could you explain that concept?

Yasuo Kuniyoshi: Yes. The neck point is a critical point in an action. In the roll-and-rise motion, there is one moment where the system transitions, and this point determines whether the action succeeds or fails. The embodiment creates this structure naturally.

Q30: It is interesting that your career began with artificial intelligence and computers, which are often thought of as disembodied systems, but your research moved toward very concrete physical embodiment. How do you reflect on that transition?

Yasuo Kuniyoshi: For me, it was just a logical outcome. I kept asking: what is behind this? Once I understood something or discovered something, I asked what was behind it, and then what was behind that. Following that chain of questions, I eventually reached the body. The questions about what is behind the brain led me to the body.

Q31: We are going to move briefly to questions specifically about humanoid robots. Before that, could you share some reflections on how you see robotics changing over the years and where you think the field is going in the next 10 or 20 years?

Yasuo Kuniyoshi: This is a very important and dramatic time because we are seeing the merger of AI and robotics, particularly humanoid robots. The evolution is happening very quickly. Finally, we may see truly intelligent robots that can perform tasks in the real world. However, there are still very difficult challenges. For example, can AI robots deal with completely open-ended and unstructured situations? Can they pursue very long-horizon tasks even under many kinds of perturbations and unexpected changes? For this, the robot needs to maintain its own purpose and understand what it is trying to achieve. In some situations, robots may even need to generate their own goals. For example, if a robot is training itself, it needs to create purposeful actions rather than random ones. I think this is one of the major challenges. The community needs to address these problems and solve them in the near future because society is placing increasing expectations on robotics.

Q32: When you say we do not have much time, what are the pressures creating that urgency? What could happen if we fail to address these problems?

Yasuo Kuniyoshi: For example, if we miss the timing, we might enter another kind of technological dark age. But more practically, society has increasing needs. In Japan and many other countries, populations are aging rapidly. There is a strong need for autonomous, general-purpose robots that can assist people. Society is waiting for these kinds of robots.

Q33: What makes a robot a humanoid? How would you define a humanoid robot?

Yasuo Kuniyoshi: A humanoid robot is a robot that can interact with the world, including humans, in a similar way to humans. If you focus on a specific type of interaction, then the body shape does not necessarily need to be completely human. For example, if the task is only desktop interaction—manipulating objects on a desk with humans—you may not need a lower body. A robot with two arms and vision capabilities might still be considered humanoid in that context. But if you consider a wider range of interactions, including whole-body tasks, dancing, and open-ended activities, then the whole-body structure becomes important. Ultimately, if we want a robot capable of complete interaction with the human world, the body shape and capabilities need to become close to humans.

Q34: That is interesting because it seems to reverse the usual way of thinking about the body. Instead of saying a humanoid needs a certain body shape, you are saying the body emerges from the type of interaction it needs to support.

Yasuo Kuniyoshi: Exactly. If you focus on a particular task, there are many possible solutions. The robot does not have to have a human-like shape. But if you widen the scope and include indefinite, open-ended interactions, then you reach the conclusion that a human-like body is necessary.

Q35: So if we imagine something like Kafka’s example of a human transformed into a cockroach—something that can still sense and move through the world but no longer has a human form—would that change the kind of embodiment it has?

Yasuo Kuniyoshi: Yes. A cockroach embodiment is a cockroach life form, a different niche. If you created a robot with a cockroach-like embodiment, it would not be a humanoid robot. A cockroach robot could probably achieve some human-like capability for a very specific task, but if you changed the focus, that equivalence would disappear. The reason we pursue humanoid robots is because we want to address complete openness and generality. In twentieth-century engineering, optimization was the central idea. You define a specific task and conditions, then design the optimal mechanism and control system. But now, in the twenty-first century, we need to move beyond that. We need systems with indefinite capabilities across many different situations. For that, we need to start from embodiment. Humans provide the reference because we are operating in a human world. Everything around us—tools, environments, social systems—is designed for humans. So humanoid embodiment is a natural starting point for designing general-purpose systems.

Q36: So having additional arms or a different body structure might seem advantageous for a specific task, but it changes the whole world of possible interactions?

Yasuo Kuniyoshi: Yes, exactly. Having four arms might sound better, but choosing a humanoid body gives a very specific meaning. By choosing a human-like body, you define a whole world of potential capabilities that are not explicitly programmed. The body has a particular structure. If you choose an octopus embodiment, then you define a completely different world.

Q37: Are there particular applications where you think the general capabilities of humanoid robots will be especially important?

Yasuo Kuniyoshi: Generality itself is the important thing. For example, if I created a company producing robots and had a technology enabling general-purpose AI humanoid robots, I could say: if you buy one, you can use it for anything. That is extremely cost-effective. Also, in disasters, these robots could enter dangerous environments without risking human lives and perform tasks similar to humans. A humanoid robot might originally be purchased for manufacturing, but then suddenly it could be assigned to disaster response or another completely different task. That flexibility is the ideal situation for future robots.

Q38: When did you first begin working on humanoid robots?

Yasuo Kuniyoshi: It was when I was at ETL, around 1995. We started a project on a whole-body humanoid robot.

Q39: What problem were you trying to solve with that project?

Yasuo Kuniyoshi: We first discussed the meaning of having a human-like body. One important situation is whole-body contact motion. If a robot touches the environment with many parts of its body, then the outer shape becomes important. Another situation is highly dynamic movement. When a robot performs dynamic whole-body motion, mass, inertia, and power become critical. The robot needs to be similar to humans in those dimensions. So we chose whole-body contact and dynamic motion as the first target for humanoid action.

Q40: How do you see the relationship between a robot’s body and its controller?

Yasuo Kuniyoshi: The robot controller should exploit the embodied properties of the humanoid robot. Typically, for example, in trajectory control or model-based point control, people tend to ignore the dynamics of the body. There are technical approaches called gravity compensation or dynamics compensation, which basically mean canceling out or eliminating the natural dynamics of the body. I do not think this is the right approach. Instead, the controller should explicitly employ and exploit the dynamic properties of the body to achieve truly embodied behavior. I think the latest reinforcement-learning-based whole-body control approaches are actually succeeding in this direction, even without an explicit notion of embodiment. They are learning to use the physical properties of the robot body rather than fighting against them.

Q41: During your career, you have witnessed the evolution of humanoid robot design and control, including the early steps toward commercialization. Can you give a brief account of this historical development?

Yasuo Kuniyoshi: I do not have a particularly unique view; I think I share a perspective with many robotics researchers who have followed humanoid robotics for many years. What is really striking is that, in recent years, we have seen the emergence of a new generation of humanoid robots. The combination of new actuator technologies and reinforcement learning has been very successful, and I think this represents a revolution. In the older generation of humanoid robotics, researchers designed robots assuming model-based control. For example, they relied on inverted pendulum models for walking control. Under this approach, the robot had to be very rigid because it needed to follow the model exactly. If there was an error in position or movement, it often meant failure. This philosophy is completely different from the latest reinforcement-learning-based robots. Actually, when I built my humanoid robot many years ago, we had a similar idea. Even 20 or 30 years ago, we discussed that rigid-body design and purely model-based control would not have a future. We believed this approach would be too limited. So our robot was deliberately designed with the lowest possible reduction ratio and very high-power motors in order to accommodate compliance and respond to external forces. Now, I think this concept was correct, but at that time the technology was not mature enough. Later, with the combination of reinforcement learning and direct-drive, low-reduction-ratio robots, learning algorithms could freely interact with the external world and learn from it. I think this is the most important change over the last 30 or 40 years, and I believe this current revolution will continue.

Q42: Do you remember any unexpected successes, failures, mishaps, or interesting anecdotes from your career?

Yasuo Kuniyoshi: Maybe this is a little different from what you are asking, but I want to mention that when I got married and had a baby, that experience changed my research direction toward developmental issues. Actually, many researchers describe similar experiences. Watching babies learn is amazing. They learn extremely quickly, adapt to situations, and develop new capabilities. Seeing this strongly motivated me as a researcher. My mind is driven by curiosity—to reveal these mysteries and try to reproduce them artificially.

Q43: Have you seen the results of your research applied in real-world situations? If so, what does that look like?

Yasuo Kuniyoshi: Our research is mainly basic research, so it has not been directly applied as a complete system. I hope that our findings about embodiment will eventually be applied to future robots, but currently the core ideas are not directly commercialized. Some technologies, such as tactile sensors that we developed, have been commercialized by companies and used by research laboratories. But beyond that, there has not yet been a large-scale practical application.

Q44: Do you have any advice for students interested in robotics?

Yasuo Kuniyoshi: My recommendation is to look at how humans, or how animals other than humans, achieve the capabilities that you want to create. That is what I did during my PhD work. I challenged extremely difficult tasks compared with the technology available at the time, but these were tasks that humans could perform easily. Of course, there are researchers who say, “Do not copy nature.” I understand that perspective. However, I would like to raise a small controversy: this idea of not copying nature is closely connected to traditional engineering philosophy, which is based on optimization. Optimization works well when you can clearly define a particular task and its conditions. But now we need to address open-ended capabilities. For these problems, we should look at how biological systems have solved the challenges of surviving in a complex world. So I recommend studying nature.

Q45: Earlier you mentioned reading all of the AI literature in the library. Would you still recommend that approach to students today?

Yasuo Kuniyoshi: I actually would recommend that kind of effort. It is very good training.
Of course, today the amount of literature is exploding, so it is practically impossible to read everything. But students should still make the effort to seriously read papers themselves, without relying too much on easy summaries or tools. Through that process, they develop their own understanding. More importantly, they develop a sense of what is important, what is interesting, and what is missing. You need to develop that instinct as a researcher.

Q46: Is there anything else that you think we should have asked, or anything important that we did not cover?

Yasuo Kuniyoshi: No, I think we covered many important things. Thank you very much for being so generous with your time and for giving me the opportunity to talk about my experiences.

Interviewer: Thank you so much. This has been wonderful. I really appreciate you sharing your experiences, your research history, and your thoughts about the future of robotics.

Yasuo Kuniyoshi: Thank you. It was my pleasure.