Professor Frank Chongwoo Park’s Lab: Robot Automation Lab

Dr. Che-Sang Park, head of the laboratory

Q1. What is the main research field of the laboratory, and what are some specific examples?  The research field as a major topic is Robotics, and among them, Robot Learning, which incorporates machine learning, is being researched intensively. Existing robots are standardized. It operates in an environment where input is constant and within control, and efficiency has already peaked. Now, in order to develop further, it must work in an atypical environment rather than a controlled environment. Unlike humans, robots have difficulty picking up objects that are not previously used for learning, and do not know information about objects. These problems cannot be solved with conventional robotics. The main purpose of Robot Learning is to solve the limitations of existing robotics through machine learning techniques.

Q2. Do you have any special equipment or facilities in your laboratory?
The equipment used in our laboratory is the ‘Frank Panda’ robot with 7 degrees of freedom. This robot ise of the representative robots used for object manipulation experiments. Research can be conducted using simulation, but according to the laboratory's philosophy that it is meaningless if it does not operate in the real environment, the algorithm or research developed in the laboratory is applied to this robot and the results are shown. The robot's main purpose is to produce demonstrations of the robot and show that it is available in real life. In the future, we plan to equipe more robot and conduct research the interaction between robots.

Q3. Apart from this, is there anything special about the lab?
The special thing about our laboratory is that we can know the direction of future research in the field of robotics more quickly. The professor is in contact with many robotics professors abroad, and they have a lot of conversations about what direction robotics will go in the future. After that, he talks to us and tell about the field of study. I think it is a great advantage in conducting research because if you can predict the future research field to some extent, you can preoccupy the so-called blue ocean in advance.

Q4. I would appreciate it if you have something to say to undergraduate students who are thinking about graduate school or graduate students who are doing research.
As a researcher, I would like to emphasize to my juniors that research is different from studying as an undergraduate. Of course, there is a high possibility that a student who did well in his studies as an undergraduate will do well in research after coming to graduate school, but it isly a possibility and not absolute. I felt it a lot. When I was an undergraduate, I got good grades in my professor's class, but when I did my research, I realized that it was different from what I had originally thought. I think this is because problem solving is the focus of studying as an undergraduate, and the answer to a problem is fixed, but research is not. Since there is no fixed answer to research, it is necessary to look at it with a different attitude than when you were an undergraduate student. In my case, when I do research, I look at previous studies and ask, ‘Has the problem been completely solved?’ Or, while experimenting, you think, ‘Why aren’t results coming out, why is this not a perfect answer?’. At the same time, a new research topic is selected and proceeded. As such, I would like to say that the direction of thinking for problem solving that I thought in undergraduate school is different from the direction of thinking for problem solving as a researcher. So, there is no need for undergraduate students to think that they are not qualified as researchers just because their undergraduate grades are low.

________________________________________

Representative researcher Seongyun Kim

Q1. Please briefly introduce the selected representative research.
When controlling a robot, it is very important to know exactly where the object being pushed or moved by the robot is located. For example, it is important to know the position of an object 1 second after it is pushed with a force of 1 Newton. However, previous studies have not been able to apply the equivariant characteristic that changes in input are directly reflected in changes in output to this physical phenomenon. This paper solved the above problem and was selected for this CoRL Conference Robot Learning conference.
Existing studies have predicted the position of an object through machine learning, but it has not been applied when it is rotated with a coordinate system. Since the data was immediately input into the model after observation in the global coordinate system of the scene where the object was manipulated, there was a problem that if there was a slight change in the manipulation situation, it would become different data. In this study, we created a model that can predict the next position of an object well even if the coordinate system is rotated. Since the manipulation situation was created as data in the relative coordinate system of the object and then inputted into the model, even if there is a change in the manipulation situation, the model has the equivariant characteristic that recognizes it as the same data. We applied this to the Frank Panda robot and it showed successful results in practice. In order to grab a flat plate the desk, the result was shown by moving the plate to the end of the desk to secure a holding space, and it was seen that if there was an obstacle around the reception desk, it was learned to push the obstacle away to grab the plate. This is a result of the robot accurately predicting the next position of the obstacle or plate.
Unlike studies in which the manipulation of an object was observed in a global coordinate system and the data was immediately input into the model and a slight change in the manipulation situation became other data, the manipulation situation was created as data in the relative coordinate system of the object and then converted into the model. Because it was entered, it has the equivariant characteristic that the model recognizes it as the same data even if there is a change.

Q2. What was the most difficult part of the research and how did you overcome it?
The most difficult thing during this research was time. The research topic was conceived two months before the submission deadline for the conference, so it was difficult to conduct the research in a short period of time. The model was coded in a hurry, and there were a lot of bugs and errors. Overcoming seems to have invested a lot of time. I really continued to research while correcting errors through debugging. There ise thing I realized through this study, that I need to systematically plan my research. I felt the need for a more systematic study because it became an inefficient study after spending time in a hurry in this study.