Interview Transcript
Q1: Hi professor. Thank you so much for your time today. Could you tell us your name and your background?
Christian Ott: Yeah. Hello, my name is Christian Ott. I am working at the University in Vienna, Austria, and I started robotics at the German Aerospace Center in Germany, close to Munich, in a time. It was about 2001 when I started to work on elastic joint robots, which are robotic systems where in the actuator a certain amount of flexibility is present, for example from the gears or other components at that time. We tackled this by integrating small torque sensors in the joints, and by this we were able to handle vibration in the joints and also to make the joints quasi back-drivable by active feedback control, which means that when you push the robot, it would move back automatically without the motor initiating this motion. So the external force is shaping the motion, not the motor itself.
Q2: Is this the first time you started working on humanoids?
Christian Ott: So humanoids, I would say I started to work on humanoids when we tried to apply these concepts first to bimanual systems, because this is also half a humanoid right now, a body with two arms. And there this torque control allows also to control internal forces. For example, when a robot grasps an object, it has a force on the right hand, force on the left hand, and with purely position-controlled devices, you would easily get internal stress, right?And by having joint torque sensors, we can implement a compliant behavior and then manipulate objects in a compliant way. Also, we can interact with the external environment like holding a handrail or similar in a compliant way. So that was my first, I would say, encounter with humanoids when we applied this to bimanual systems.But later, it was about in 2009 we decided to try similar concepts also on bipedal systems. So we built up with the same actuator concept bipedal systems. And our idea was that the humanoid robot with legs makes a contact at each time it makes a footstep, right? And then this compliant controller, which we used before on the manipulation side, should also be useful for locomotion. So this was our concept and we built up a first bipedal system. First it was a bipedal system. Later in 2013 it was extended to a full humanoid with upper body. And it was interesting to see for us that this kind of concept could compete quite well with the current technology at that time, which were mostly position-controlled humanoid robots. Because at that time compliance was implemented mainly by force torque sensors in the ankles in combination with position-controlled actuators. So the actuator itself is stiff, but it makes the compliant whole-body behavior with these force torque sensors in the ankle. And our idea was to use these torque-controlled actuators instead. And this has quite some advantages because then any force anywhere on the body would also be considered in this compliant control. So the robot can react in a compliant way if you push at the hip, at the shoulder, or anywhere. And not only — you don’t only have to measure all forces individually at the end effectors, but you can react to disturbances and other interaction forces in a compliant way. And first, we showed that this new activation concept at that time also works for balancing on uneven terrain or on some, let’s say, movable surfaces, where the idea was more to control the ground interaction forces in such a way that our center of mass would stay well balanced within the support area. And this concept extended nicely also to locomotion control. Because locomotion is mainly shifting the center of mass from step to step. Later, we found that the center of mass is maybe not the best quantity to control. Instead, people went to the diverging component of motion, which is a combination of the position and the velocity of the center of mass. So it’s a more dynamic quantity. And also this dynamic quantity can be well controlled with these torque-controlled actuators by implementing whole-body control algorithms where the interaction forces between the feet and the ground are mapped basically to the forces activated on the center of mass of the system. Yeah. So this was, I would say, our first attempt to implement torque-based control on a biped and humanoid robot. And we found also not only for locomotion but in particular for multi-contact interaction, this was very useful. Because in multi-contact interaction, if the robot walks and holds a handrail or it has to stabilize itself with two arms with some environment, then by having an intrinsically compliant actuator control, we can handle the internal forces easily. With a position-controlled device, you have to have a very precise position control in order not to build up too much stress in between the contact points. And the compliance control allowed us to handle these multi-contact situations in a very nice and very robust way.
Q3: What are the key publications which you look at if you want to read about the things just mentioned?
Christian Ott: Okay, so I like my earlier publications maybe more, where we built the foundations on this torque control. So these were publications in the Transactions on Robotics and in the International Journal of Robotics Research. This was about 2007, 2008, where we put the foundations on compliance control using joint torque sensors. And later, of course, when we applied this to the humanoid robots, we had again publications in the Transactions on Robotics, for example on the DCM-based locomotion. This was coming in 2013. And of course there are many follow-up works which build up on these concepts.
Q4: What was difficult in early days when you started working on humanoids?
Christian Ott: So I think what was really difficult was to have the very first basic control behavior running on the system. Because in whole-body control, if you are in torque-based control, if your algorithm does something wrong, the interaction forces with the environment are wrong and the results can be very dramatic. The robot would not only fall down right or left in a slow way, but it would kind of explode in one sense. And at the time, we had to be very careful with the robotic system because we didn’t want to break it, right? And when we make every step, also, we had to be careful not to hit the ground in a too impactful way, otherwise the torque sensors would be affected or even the gears could break if the impact is higher. So I think this technology of torque joint-based actuation systems is very nice for compliance control, also for human-robot interaction, but it has some disadvantages, I would say, in terms of impact robustness. And if we compare now with humanoid robots which are currently built, we see quite robust, impactful-robust hardware designs which allow much more aggressive algorithms to be implemented. And of course, at the university, we have to be a bit careful because if we break our system, we have to send it back for repair or try to repair ourselves. And this would put a strong pause in our research. So impact-robust hardware, I think, is quite important aspect in our days’ research on humanoid robotics, because you can implement more aggressive and more dynamic behaviors without the risk of damaging the robot so easily.
Q5: So speaking of software and hardware, how do you see the relationship between robot’s body and its controller?
Christian Ott: I think it’s very closely related, because the way we control the robotic systems are based on the internal dynamics, right? And then we have either torque control or position-controlled actuators. And they behave quite differently, right? If you have a back-drivable system, you have kind of zero impedance. It moves very easily. While with the position-controlled device, you have a very stiff system. You have to make it compliant by active feedback control. And so the intrinsic behavior is very different. And that also affects the control algorithms. The position control robot you would control with a different algorithm than a torque-based algorithm robot. And also you would control it differently, maybe to a more robot which is nowadays direct drive or quasi-direct-drive actuators. So I think there’s a very close relationship between how the controllers are developed and how the robots are built. So I think it’s really a co-development nowadays. We see also a lot of AI-based algorithms, and in the media, I feel sometimes the performance, the very nice performance of modern robotic systems, is projected in a strong way to these AI algorithms. But in my personal opinion, this only works because it has been really a co-development of robust hardware and more aggressive algorithms.
Q6: Has this sort of core evolution affected the way you approach the design of a humanoid robot?
Christian Ott: Yeah, I think it affected a lot the way we design humanoids. So I studied for quite some time also elastic humanoid robotic systems, because then it is possible to catch the impacts, not by the active controller, but by the mechanical components. Then the robot would be impact-robust by the mechanics. Of course, the control becomes more complex because in an elastic robot it is also dangerous to have vibrations at some places. You don’t want to have them, so the whole-body control becomes more complex. On the other hand, the physical robustness is increased. And I think compliant design in humanoid robots is also one of the areas where there’s a lot of potential for future research.
Q7: So you’ve been in this sort of humanoid robotics world for quite some time. Could you tell me how it has evolved when you started to this day?
Christian Ott: I think I feel I started late to work on humanoids because I think the foundations of humanoid robotics really have been made in the early 2000 years. Maybe in the 90s, people started. You can go a bit earlier even if you want to look. But a lot of progress has been made around the development of assembled HRP-2 robots. Also, there was a big trend in Japan to have research on humanoid robotics. It was more research-oriented. I think it was not that time that companies had such big interest. But research-wise, I think Japan was very strong then. And now, recently, we see humanoid robotics driven by commercial interests. Also everywhere in the world, mainly the US and China are very strong now, but also in Europe we have some activities and we hope also to make our contributions there.
Q8: Who are the biggest contributors nowadays in humanoid robotics?
Christian Ott: So I think the biggest contributors on the financial side are companies, big tech companies, who want to invest in this business for the future. I think there is — yeah, it’s not clear for everyone what will be the killer application and what will be really the market opener. I mean, there are some companies which have very clear use cases. On the other hand, I think for most of them it’s investment into the future, that when the market opens, people want to be in the position to bring out the systems right away and don’t have to do very long basic developments before starting to roll out the systems.
Q9: Are there any labs or research groups that people should check out apart from your lab when it comes to humanoid robots?
Christian Ott: Oh yeah, I think there are many labs in the world. So of course AIST in Japan has a very big traditional lab on humanoid robotics. But also in the US there are many big labs.
Q10: Is there anything else you’d like to add about your research?
Christian Ott: I think one aspect which was not addressed so far was that what we see in humanoids now also is the trend of bringing research results into commercialization. So we see that the algorithms being developed are really used by companies. And that is something which affected my research in the past, I would say, because I started to work on elastic joint torque-based robotic systems, which then was commercialized later to the KUKA lightweight robot. And then seeing how these algorithms are used really in practice and that they are relevant to companies, I think was very interesting. And I think the same is now happening also in humanoid robotics, that results from research are really taken into practice.
Q11: That’s very exciting. Yeah, that’s all my questions. Thank you so much.
Christian Ott: Thank you.