研究最前線 No.63

徳丸 正孝教授


Research and development of an intelligent model for emotion generation



Research of Kansei* Robot behavior

Creating a robot that can empathize with humans
* Kansei : Japanese word which has various meaning related with human sense, feeling, affection and emotion.


徳丸 正孝教授

Faculty of Engineering Science

Professor Masataka Tokumaru


Robots are slowly becoming a part of our daily lives, whether serving customers at hotel front desks or living with people as part of a family. However, there are fears that artificial intelligence (AI) may negate the need for human employees and with further development, control human society. How can AI and robots coexist and prosper alongside humans? Professor Masataka Tokumaru of the Faculty of Engineering Science is researching to find ways in which robots and humans can sympathetically communicate with one another.






Developing emotional robots

I understand that you research AI and robotics. Can you tell us more about your research?

Thanks to the evolution of AI, robots are now able to distinguish the smallest details of human emotion based on the differences in subtle facial expressions and voices of humans. We now want robots to be able to express emotions like humans. To achieve this, we are designing an emotion generation model through which robots can express happiness or sadness depending on various external stimuli, similar to humans.
Will robots be able to develop empathy by sharing similar emotions through conversations with humans? It will be interesting to assign short-tempered or optimistic characteristics to individual robots and have them communicate with humans.

What stage of research are you currently in, and what type of experiments are you conducting?

For the emotion generation model, we have been trying to achieve emotional expression by the interaction of humans and robots through repeated communication, similar to the developmental model of educational psychology through which human children's emotions gradually mature.
Presently, robots as hardware are not able to express emotions effectively. We are designing several robots in a virtual world as fictional characters using 3D computer graphics (CG). As humans communicate with these robots, they learn how to move; this approach is similar to training animals. As each computer graphics robot has a unique character, we varied the training method for each robot and designed the experiment so that the robots and humans would both be pleased when the robot is able to perform its actions well. Using this model, we observed the behavior of robots.
In parallel with emotion generation, we have also begun conducting experiments to observe whether robots and humans could anticipate each other's intentions during cooperation.


The image shows a behavioral training experiment for a slime robot inspired by dog training. The robot learns which direction to move in response to the glow of the tiles, but since it is equipped with a model of character and emotion, these parameters change the effi ciency of learning (cleverness) and emotional expression (honesty,emotional ups and downs, etc.). By understanding the robot's character and training it, the participants are expected to become friendly with it. (Currently, the experiment is being prepared and will be carried outperformed in the future.)


どのような実験ですか? ロボットも空気を読めるのでしょうか?


Can robots guess intentions and act accordingly?

Can you describe this experiment? Were robots able to guess the intentions?

A student in the lab suggested that it would be interesting to conduct an experiment in which robots and humans maintain an ideal distance from each other. In this scenario, it was difficult to discern other's emotions from conversations and guess intention; however, with physical distance, the stimulus applied to the robots was less complicated and data were obtained more easily.
First, we set up four CG robots in a virtual world. Three robots were controlled by humans, and the fourth robot was set up to move according to its own judgment. All four CG robots then played a game of "catch." The robots controlled by humans would look at each other while separating and stop at some distance from one another. The robot acting on its own was able to observe how far away the three other bodies had moved from one another, and it was able to relocate and stop at the same distance before starting the game.
This experiment was completed virtually. We are currently planning to conduct similar experiments in real life with a human and a robot to observe whether they can establish the correct distance while congregating. The robot will be able to move back and forth and left and right, and it will be equipped with virtual reality sensors to detect the positions of the human and the robot, and each movement will be reflected within the virtual environment.
However, we are struggling with the aspect of setting up a realistic situation for the robot; for example, "We are discussing this painting in a museum, so let's gather around the painting," or "Let's join a circle of people having a discussion." When people congregate, they tend to maintain some space between themselves and others. This uniquely human approach to movement is, in fact, difficult to achieve but it is what I wish to achieve with robots. This is what I mean by "guessing the intention."


The image shows an experiment of a robot learning "just the right distance" through a group normative learning model. One robot is immobile, two users operate one robot each, and the remaining robot operates autonomously by a learning model.The two user-operated robots and the autonomous robot move away from each other "as if they were playing catch," and the autonomous robot learns in real time while checking the distance between them. This motion is recorded on video, and subjects are surveyed on their impressions of whether each robot is leading or following the other in terms of distance.




Robots as partners

How do you envision a future with robots around you?

Developing a robot as an efficient and convenient tool is one approach. However, I think it would be entertaining to have a robot that may not be very useful as such but is simply an interesting creature.
If I continue to train the AI so that my pattern of behavior in every aspect of my life is correct, the AI will do and think completely like me. Accordingly, it may be able to complete tasks that I do not wish to do myself. That is not where my interest lies, however.
If a robot starts talking to me and caring about me, I am sure I will become attached to it. I believe it will be enjoyable to work alongside it. The ideal goal is to create a warm and caring relationship with AI, rather than simply pursuing efficiency and convenience. I believe one of the key aspects of realizing this is emotion.


The image shows a robot built by students in the research group for the experiment.




How will robots change humans?

What have you learned in your research on AI and robots?

I have been developing AI with emotions and with the research's progress, my interest has shifted toward observing how human behavior changes while interacting with AI. When I talk to other researchers, I do not find this surprising because the research also leads to studying how human life responds to changes within social systems. I believe that a degree of perspective had been lacking when the Internet was first created.
At many instances, AI is discussed as something to be feared, that will take away people's jobs or control humanity, as it is portrayed in films. This is a future possibility; however, if this happens, it will be important to address how we interact and engage with AI. I would like to be able to metaphorically shake its hand and suggest, "Let's get along."