November 27, 2017      

For decades, people have dreamed of an effective humanoid robot, but as the DARPA Robotics Challenges demonstrated, there is much work to do to realize that dream. Joanne Pransky, associate editor of Industrial Robot, recently chatted with Dr. Jun Ho Oh, professor of mechanical engineering at the Korea Advanced Institute of Science and Technology (KAIST) and director of KAIST’s Hubolab.

Dr. Oh obtained his B.S. and M.S. in mechanical engineering from Yonsei University. He worked at the Korea Atomic Energy Research Institute before getting his Ph.D. from the University of California Berkeley, where he also conducted postgraduate research.

Oh has been a professor of mechanical engineering at KAIST since 1985. He also serves as a robotics policy consultant for the Korean Ministry of Commerce, Industry, and Energy and is a member of the Korea Academy of Science and Technology.

In the early 2000s, Oh built the KHR1 to compete with Japan’s humanoid robots. This led to seven successive versions of the bipedal humanoid robot and the founding of the Robot for Artificial Intelligence and Boundless Walking (Rainbow) Co.

In 2015, South Korea’s DRC-HUBO+ humanoid robot beat 23 teams from six nations to win the $2 million DARPA Robotics Challenge (DRC).

Dr. Oh discussed the challenges around humanoid robots and efforts to commercialize them.

This interview is available for free to Robotics Business Review readers until Dec. 31, 2017. Here’s a preview:

Pransky: How did you first come up with the idea of developing a biped humanoid robot?

Oh: My motivation came from Honda’s Asimo and Kawada’s HRP-4 in Japan. In 2001, I was curious if I could make something similar, though I did not have a specific vision.

The problem was that I didn’t have any special qualifications, nor were there the necessary funds available. However, I got a little money from my colleagues, and I built my first bipedal robot named KHR1.

I was motivated to build my own humanoid robot in a very short period to compete with Japan’s humanoids, and I completed HUBO1 in my third year, in 2004.

Pransky: Can you please describe the evolution of HUBO, your funding sources, and why you founded Rainbow Robotics?

Oh: For 2002 and 2003, we were given a small amount of money by KAIST, and from 2004 to 2009, the Korean government supported the project.

The average budget per year was about U.S. $500,000 per year for six years, for a total of $3 million for the development of HUBO1 to HUBO2.

Around 2009, we produced HUBO2, which was one of the lightest humanoid robots at the time. It weighed only 42 kilograms, but the performance of HUBO2 did not quite match the performance of Honda’s Asimo because we did not have enough funds for further development. Instead, we were limited to off-the-shelf motors, reduction gears, etc.

South Korea's KAIST has been working on humanoid robots for years.

KAIST humanoid robot research timeline

In 2010, we got two simultaneous inquiries — one from universities in the U.S. that wanted six HUBO2s, and the other from the Singapore government, which wanted two HUBO2s.

It was impossible for me to produce eight HUBO2 robots as well as maintain them for an additional three or four years, so I consulted my university presidents and other colleagues, and we concluded that it would be better to establish a company to manage these opportunities.

As a result, we established the enterprise Rainbow Robotics in 2011 with the purpose of producing and supporting HUBO robots. After the establishment of Rainbow Robotics, we received more orders, and we delivered in total somewhere around 20 HUBO2s all over the world.

Pransky: Are you able to discuss Rainbow Robotics’ long-term mission or objectives?

Oh: We are currently working with a major Korean venture capitalist, which has already given us some initial funds and will be making significantly larger investments in Rainbow Robotics.

Pransky: Why did Rainbow Robotics refuse an offer from Google to be purchased in 2013?

Oh: Google’s Andy Rubin visited my KAIST lab for a period of two days. He and I spent a great deal of time discussing the future markets of robotics and development.

Andy showed me his plan of Google’s mergers and acquisitions in different robot companies all over the world. He asked me if Rainbow Robotics could move to the U.S., and he also asked me to send some IP [intellectual property] or software to Google.

Andy did not state a specific amount of money, but he asked me with general intention. At the time, I was a vice president of KAIST, and Rainbow Company was just starting out. I simply answered, “I don’t think we’re quite ready.”

Six months later, Google acquired eight robotics companies, but Andy did not follow up to our discussion. I ran into him at the DARPA Robotics Challenge trials a month later in December 2013, and I asked him why he purchased the other eight companies but not Rainbow Robotics.

Andy told me, “I offered it to you, and you refused; that’s why.” He also added, “I’m open if you change your mind. Please call me.”

Pransky: Did you come up with the addition of the wheels and the knees while you were redesigning your humanoid robot for the DRC’s eight tasks? Could you please describe how the system works? Are the wheels in both the knees and feet powered?

Humanoid robot HUBO1 during the 2015 DARPA Robotic Challenge.

Team KAIST’s DRC-HUBO1, winner of the 2015 DARPA Robotic Challenge, displays its agile kneeling.

Oh: Yes, that idea came from our experience at the DARPA trials. I observed that most of the teams in the trials struggled with bipedal walking. Even though their performance may have been good for the demonstration, the advanced algorithms used in the experimental setup were not going to be robust enough to maintain strong bipedal walking in actual practice.

Thus, I decided to develop a new type of mobility. First, we applied transformations to quadrupedal walking and bipedal walking in my lab. In the stable walking position, I decided to use a quadruped walking with arms. The robot stood up and performed some simulated DRC tasks.

Quadrupedal walking is very difficult with two arms because the arms are not strong enough to support its entire body weight.

When this approach failed, we devised a type of wheel on the knee or thigh for long-distance movement. We tried three or four different versions, including caterpillar-type orbital wheels, and finally, we decided to put one wheel on each knee and an idle wheel on the toe of the foot.

In the next test, we tried to make the robot transform from the biped mode to the wheel mode. The movement from the kneel function to the upright position and vice versa was not that difficult and was very successful.

Pransky: Is the addition of the wheels in the knees an important feature and an example of a good design that is better than numerical computation?

Oh: Yes, DRC HUBO1 is a good example of a good design. In the future, I think all robots should have different kinds of mobility modes. For example, a drone with some kind of propeller may need an arm for some kind of landing or require the ability to swim underwater.

A patrol robot may need more mobility other than wheels to, say, climb up stairs, or [it] may need arms or legs for other situations.

Specs for the DRC-HUBO1 humanoid robot

DRC-HUBO1 specifications (Click here to enlarge.)

Pransky: Have you sold more humanoid robots since the DRC and to whom?

Oh: Through Rainbow Robotics, we are selling many DRC HUBO1s to research labs around the world: a U.S. National Lab has already purchased three or four DRC Hubos1; Korea Atomic Research Institute has purchased one, and we are reparing three more for other Korean research institutions.

We have several inquiries from Europe and some other Asian countries. There are limited choices available in today’s marketplace for a robust humanoid platform to perform sophisticated research requirements.

DRC HUBO1 won the DRC, and it is a proven product. It was one of just a few robots that solved all eight tasks with enough power, good mobility and a completed development environment.

Rainbow Robotics provides research institutions with all the simulators, algorithms, and all the development details and service necessary, so the labs can replicate similar tasks for their own purposes.

More on Humanoid Robots:

Pransky: If you are comparing DRC HUBO1 with a person, how would you rate the performance of the humanoid robot in percentage terms? For example, do you think DRC HUBO1 is 10% as good as a person, or what percent?

Oh: It is very hard to say. I should say frankly, less than 20%.

Pransky: And when do you expect DRC HUBO1 to be 50% or 80% as capable as a person?

Oh: It will take a very long time because, thus far, a humanoid robot cannot replace humans one by one because its functions are too inferior to humans. Our first-place robot in the DRC took about 40 minutes to complete all the missions. The same tasks can be done by a human in less than two minutes.

Furthermore, the robustness success rate by DRC HUBO1 may not be 100% successful. DRC HUBO1 may repeat the same success rate of finishing in 40 minutes only 70% or 80% of the time. A human, however, will take two minutes, and the success rate is almost 100%. A robot replacing a human at a task will be more than 20 times slower with a success rate incomparable to the human.

Robots are needed in the field in very specific cases such as in space missions or in nuclear power plants or other harsh environments where a human cannot go. We can use the DRC HUBO1 for research purposes, but in regular situations, it is very hard to replace the human with a 1:1 rate. It will take time.

Pransky: What is the greatest lesson or the biggest mistake that you’ve made in your career?

Oh: The most significant lesson for me was the DRC trials. From the lesson of ending up in the bottom half came the development of the DRC HUBO1 that won first prize. Without having that lesson, it was impossible to have such a great result at the finals.

Click here to read the entire interview.

More Essential Interviews: