Working in scientific research tends to require manual dexterity and long working hours. Project manager (PM) Harada Kanako believes that training AI robots to become “clever, skillful scientists” will remove these constraints, achieving a future where anyone who wants to become a scientist can do so. She is leading a project titled “Co-evolution of Human and AI-Robots to Expand Science Frontiers.”

What are these “AI robot scientists” that PM Harada envisages? How will this become a reality, and what will happen to scientific research when it does? 
Ikegami Akira, a journalist with a strong interest in science and technology, who is well known for explaining current affairs in an accessible way from a unique perspective, probes the topic with PM Harada.

AI robot scientists that get more clever as they repeat experiments

Ikegami: What is your vision for the year 2050?

Harada: Our vision is to create AI robot scientists that will act as new eyes and hands for human scientists, to explore science together. Robots can work in environments that are dangerous for humans, like being exposed to unknown viruses or toxic gases, and they can also perform extremely delicate operations that human hands are incapable of. This is already helpful to scientists, but by combining these robot “bodies” with AI “brains”, we want to create robots that can act autonomously.

Ikegami: Making robots “autonomous” is important, isn’t it.

Harada: Yes. Autonomy is completely different from automation.
In factory automation, robots are used to perform tasks with no variation. Once a human prepares the initial program, the robot will repeat the same action tens of thousands of times. But when conducting experiments in the life sciences, there are slight individual variations in the target organism being investigated. Repeating the same action is no good in this case, and it would be impossible for a human to teach the robot every action in advance. The robot needs to think for itself and become more clever and skillful as the experiment is repeated, gradually becoming better at what it is doing. This is what we mean by autonomous: it’s like becoming independent.

Ikegami: Is this made possible by deep learning?

Harada: It is made possible by AI using deep learning.
Let’s explain by looking at actual equipment. This is the robot platform that we are currently developing, with the aim of performing delicate experimental actions for various biological models. To show you in an accessible way how the robot can respond to slight variations in the target object, we are now watching it handle eggs of various sizes, just scraping off the eggshell without damaging the membrane.

A human would be able to scrape off the shell using an appropriate action for each egg, large or small, without measuring its properties. But to get a conventional robot to perform this task, it would have to measure the shape of the egg and the thickness of the shell each time. Also, over the course of performing the experiment several times, a human would gradually acquire intuition or tricks like “if the sound of the drill changes, it’s about to break through” or “when the white part becomes visible, it’s getting close to the membrane.” These kind of intuitions or tricks are difficult to express in words or numbers, and cannot be taught to a robot using the conventional approach.

With this robot platform, the robot arm is first teleoperated by a human so that the robot can experience the slight individual variations in eggs. The sensory data from this operation – images, sound, force and so on – are used to train the AI, which is the robot’s brain. This enables the robot to perform actions to suit each individual egg, even if there are variations in shape and shell thickness.

Ikegami: I see. The robot gets more clever by moving its hands and using its senses.

Harada: Yes, that’s right. Human brains are connected to our bodies from birth, so a human gets more clever and more skillful through a process of trial and error: using the body to try out thoughts from the brain, then sending the results back to the brain. We are getting robots to do the same thing.
But this kind of robot cannot be made by one person, so we are working on research and development based on “convergence knowledge” (*1) – integrating knowledge from experts in mechanical engineering, AI, mathematics, and the scientists who will use these robots. Rather than just simply putting researchers from different areas together, we have formed the group by deliberately selecting people from fields that can stimulate each other.

Ikegami: That’s the role of the Project Manager, right?

Harada: Yes. I think my experience of having worked in different places is useful in this management role.

Opening up opportunities for anyone to become a scientist

Ikegami: I teach at the Tokyo Institute of Technology. Even if I leave my laboratory around 1am, there are still lots of students around the campus. For researchers in the sciences, long working hours are a fact of life. Your research will ultimately lead to changes in the way researchers work, won’t it?
Improving harsh workplaces will eventually attract more people to the job, and if it can help to create an environment where it is possible to balance research and childcare, the number of female scientists should increase, too.

Harada: You’re right.

Ikegami: Dr. Yamanaka Shinya, winner of the Nobel Prize for Physiology or Medicine, is said to have become a researcher instead of a clinician because he was not very skilled at surgery. If these kind of robots had been around in those days, Dr. Yamanaka might have carried on to become a clinician, rather than researching iPS cells.

Harada: That’s true. There are lots of people who want to become scientists but give up because they are too clumsy. These robots will mean such people don’t have to give up on their dreams. In future, I think artists could be involved in research, not just doctors and scientists.

Ikegami: I see. That’s possible.

Harada: Somebody who knows nothing about science but has an amazing flash of inspiration could suggest “try this” to a robot, and the robot could try it out. I hope we can make this future a reality – a future where people can go beyond their own physical and intellectual limitations to pursue their dreams.

Ikegami: It would be possible to put a robot on the moon and control it from Earth, wouldn’t it?

Harada: Of course. But we want to enable the robot to think for itself and work on the moon without being controlled remotely by humans.

To explain a bit further, cooking involves two steps: thinking of a menu, and then making the dishes. Scientific experiments involve two steps, too: deciding what experiment to do, and then conducting that experiment. With the robot arm that we have just seen, we are using AI to make it more skillful at cooking, but we are trying to make the AI more clever at thinking of what dishes to cook, too. If the AI is trained for both of these aspects, even in an unknown environment with no prior information available, without detailed instructions from a human, the robot should be able to decide what kind of experiment to do, and perform it skillfully.

What’s more, our aim is to create robots that can carry out research together with human scientists, so we want the robot to be able to explain what it is doing in a way that the human can understand.

Towards the practical application of AI robot scientists

Ikegami: That’s true autonomy, isn’t it. But there are concerns that autonomous robots could go out of human control. For example, if a human doctor makes a mistake during surgery, the doctor will be held responsible, but what happens if an autonomous robot does something wrong? Is there a discussion going on about how to ensure ethics in such cases?

Harada: Of course this is being discussed. ELSI (*2) researchers are involved in all Moonshot projects, not just our project. They have ongoing discussions with the R&D researchers about what kind of issues could occur when an emerging technology is introduced to society, and how to solve these issues.

Personally, I do not believe it is necessarily a good idea to limit the functionality of robots. For example, in cases where conducting an experiment could adversely affect the scientist’s health, we would choose a robot even if there is a possibility of failure. That’s why we need to research how to establish rules that will be best for humankind.

As autonomous robots are gradually introduced into society, I think various issues will arise, beyond the range of possible ethical issues that have been anticipated. For each issue that arises, ELSI researchers and R&D researchers will work together to establish rules for the use of robots. I think it is important to head in this direction.

Ikegami: I’m relieved to hear that. But some people are worried that if robots become autonomous, they will take away jobs from humans. What will happen in scientific research jobs?

Harada: I think the important point is what kind of robots we make. We are taking an approach of making robots to do things that humans cannot or don’t want to do, not robots to replace humans. Humans will do what they want to do, as they have done until now, so I don’t think the jobs that humans want to do will disappear.

Ikegami: So that means it will be OK?

Harada: It means “we will make robots such that it will be OK.” We are not in a position to worry about “what will happen to AI” or “what will happen to robots.” We are in a position to determine how to make AI and robots, so we will make robots such that the jobs that humans want to do will not disappear.

Ikegami: I see. There is a famous saying by Peter Drucker: “The best way to predict the future is to create it.” That’s what you’re saying, isn’t it.

Harada: Yes, that’s right.

Ikegami: In that case, it’s a question of imagination, isn’t it. It’s not enough to simply look at what’s in front of you and say “it would be convenient to have something like this.”

Harada: All research and development within the Moonshot Program is based on ambitious visions for 2050. But what we can envision now might be different to what we can envision in ten years’ time. We need to keep thinking about the vision itself as we go along, and flexibly adapt our research and development work.

AI robots will make humans evolve

Ikegami: Another concern is that if we get these incredible robots to do things that humans can’t or don’t want to do, then humans might come to rely on the robots and lose our own abilities.

Harada: Certainly, human abilities might decline in some ways, but I believe they will evolve in other ways. For example, I think many people’s online communication skills have improved since the COVID-19 pandemic. The human brain is very flexible or adaptable in that way.

Ikegami: I got it. By interacting with robots that do things we cannot do, new human abilities will develop and grow. You mean it will produce a kind of “chemical reaction”. That’s very interesting.

Harada: That’s exactly what the word “co-evolution” means in the title of our project. We believe that as robots evolve, humans are sure to evolve too. We may evolve by getting inspiration from robots, and we may evolve as robots develop and the role of humans changes. It definitely won’t be a case of robots developing and humans staying the same.

Ikegami: I see. That’s why you mentioned that robots need to have the ability to explain. If we cannot understand what a robot is thinking, like a black box, then co-evolution will not be possible. That’s why it’s necessary to visualize the robot’s thoughts.
Your ideas are really fascinating. I feel that this project has endless possibilities.

Harada: Thank you. We have lots of dreams, and we will do our best to make them a reality.

Written by Aoyama Seiko
Photos by Mori Takahiro

Related information

■About Moonshot Research and Development Program

■Moonshot Goal 3
Realization of AI robots that autonomously learn, adapt to their environment, evolve in intelligence and act alongside human beings, by 2050.

■Goal 3 R&D Projects
Co-evolution of Human and AI-Robots to Expand Science Frontiers