Rolf Pfeifer

Interview with Prof. Rolf Pfeifer

Director, Artificial Intelligence Laboratory
Department of Informatics, University of Zurich, Switzerland
Personal website

What do you think cognitive systems are?

I think that is a good question. There is a definition of cognition by Colette Maloney which I like very much. She is saying that cognition is a way of overcoming uncertainty accompanying interaction in the world that is not completely pre-specified. I think the point about this is that we live in contrast to an industrial factory environment for example, we live in the real world and I think the real world has characteristics which have an intrinsic uncertainty.

We only have very limited information about the environment. Things are just happening; it is a non-linear dynamical system so it is very hard to really predict what is going to happen. But we have to live in this world and biological systems are well equipped to deal with this kind of uncertainty. I think cognitive systems are or should be systems that are capable of dealing with this kind of uncertainty. 

What is your area of research within Cognitive Systems?

We are working in what you would call bio-inspired robotics. We are looking in particular at dynamics, a lot of locomotion, biped locomotion, quadruped locomotion, underwater locomotion and we are also looking at modular robotics and then and we recently a couple of years ago started working on assistive robotics like prosthetics and general assistive technologies (i.e. technologies to support body functions, in particular for elderly people, or people with amputations). 

Why did you become a researcher?

It was not a deliberate decision and it just happened to be that way. When I was working at ETH, the Swiss Federal Institute of Technology in Cybernetics we organised a series of lectures and one of the people that we invited was a clinical psychologist. He said he was looking for someone who would work on computer simulation of dreams. So he had funding from the Swiss national science foundation on a project on computer simulation of dreams and then he asked me whether I would be interested.

He was talking about a computer simulation of neurotic defense processes and I did not understand a word and I told him I do not know anything about psychology. Well, he said it does not matter, I need someone who can do simulations and I had as a physicist, studying mathematics and physics, I had been working on simulations of neutron physics problems. Well he said it's fine if you know about simulations, you do not need to know anything about psychology and then I started working for him. That is basically how I got dragged into this so it was not a deliberate decision.

How did you get into Cognitive System research?

At the time this was a very classical approach, it was very classical cognitivistic, symbolic artificial intelligence, and because it did not work out we then got involved with something completely different.

Where did you study and what subjects did you study?

I studied in Zurich at the Swiss federal institute of technology, ETH. I studied physics and mathematics at ETH and then I did my PhD in computer science at ETH, but with the topic of computer simulation of cognitive processes.

Can you describe briefly how you are doing what you do?

We are applying what you would call a synthetic methodology. Basically we have a phenomenon of interest which could be anything and in our case it is locomotion we do a lot of. We are then building a system that at least mimics certain aspects of this behaviour that we are interested in. Because we are interested in embodiment and embodied intelligence, intelligence that has physical instantiation, of course our systems of choice are robots. So we build robots, typically biologically inspired robots, to investigate how the biological system functions. But then also to get artificial system for this particular behaviour or functionality. 

What are the techniques used in your research?

I think one of the important points is that we focus on, because of this interest in embodiment, is morphological and material properties. So it is not only the control that we are interested in but we are also interested in how the control collaborates with the morphological and material properties of the system. So we devote a lot of attention to the design of the morphology and we look at materials. We look at for example artificial muscles and deformable tissue and it turns out that often the morphological material characteristics of the system can take over part of the functionality that you would normally think have to be done by control.

Can you tell me why they are important?

I guess there are two main issues there for why I think this is important. One is that it leads to interesting and novel applications as we now do in assistive robotics and prosthetics. We also work on rehabilitation devices and there are interesting applications there. And the second equally important point is that it provides us with a completely new way of thinking about ourselves and thinking about the world.

What are the major implications of your work?

The implications of embodiment I think are really strong, and of course when you have an embodied system you need to deal with issues such as energy and physical forces. But the even more interesting point is the interaction of physical processes with information processing of the brain. For example, when you are grasping a cup you are not only grasping the cup, but you are also through your active interaction with the environment generating sensory stimulation in the hand and on your finger tips. It can also be shown that this sensory stimulation contains correlations, within the haptics channels, but also when you are grasping and lifting a cup you are generating proprioceptive sensory stimuli simulation in the arm.

Then also because you are bringing the object into the range of the visual field you are also generating visual stimulation. Then you can start forming cross-modal associations and because the sensory stimulation already contains correlations it is relatively easy for the brain to perform these cross modal associations. Then the brain can also learn how to make predictions, so over time you learn just by looking at a cup how it would feel when you actually grasp it. There are some of the really important implications of embodiment.

What skills do you think are most important to a Cognitive Systems researcher?

That is hard to say. When we hire people I do not care at all where they were coming from. I think they have to have an interest in this. Maybe it is good if they like to build things because we are using the synthetic methodology, so for us because we are interested in robotics it is if course nice when they have skills of a mechanical engineer, electronic engineer, and programming skills. But these are skills that can be learned. We have people from all kinds of backgrounds, from molecular biology, computational neuroscience, physics, I think I have about three physicists, one media scientist, one industrial designer, mechanical engineer, electronic engineer, and a couple of computer scientists and a mathematician.

What do you think is most satisfying about Cognitive Systems research?

One of the things I really like about synthetic methodology is that not only do you learn a lot about natural systems, but you also basically have a behaving system, a system that behaves in the real world and actually moves around. I think that is very nice. We also have a lot of people from television for example and they really like what we have. Because normally what they get is only stuff on computer screens and then they film it and show it again on a television screen. But when they come here to our laboratory they see actual physical things in the real world. It is very satisfying to have something that actually moves in the real world. But I mean also the kind of insights we are gaining because we are building these physical systems. It is just amazing what you learn about nature and about the real world. It is just fascinating.

What do you consider is the most challenging about being a Cognitive Systems researcher?

As I said before going back to Colette Maloney’s definition of cognition, a way of overcoming uncertainty accompanying interaction with the world that is not completely pre-specified, I think that is the big challenge. What I see often happening is that people resort to very classical methods; you know symbol based AI, Bayesian methods. I think Bayesian methods are probably a step in the right direction, but now what we are trying to do is to not get into the trap of classical artificial and symbol based artificial intelligence. There the problem is that you basically program into the system how you think the system should actually function.

So I think a big challenge is to really build systems without programming everything into the system. They definitely have to be learning systems and we can also use evolutionary methods. I think we should refrain from trying to define everything into our system and that is a really big challenge. Another big challenge is to do with the exploitation of the real world of morphological and material properties. I think that if you really want to have systems that can function well, in an uncertain environment, you need to deal with the issues of embodiment. 

What do you think are the main challenges for the future?

I do think the challenges are to build systems that basically do what you want them to do without pre-specifying everything. If you pre-specify everything, then the system is not going to be adaptive. I think the system should be adaptive and I think that is only possible if you equip them with adaptive mechanism so that they are not just programmed in that sense.

I also think it is really about the interplay between the embodied system and the control. Mostly people have a tendency to focus on the control. They basically think we have humanoid robots that are given and then they focus on the control. I think it is really about a task distribution between morphology, material properties, interaction with the environment, and control.

There are several discussions or debates associated with Cognitive Systems research.

Could you mention issues relating to your work?

Can you outline the arguments of the opposing sides of the debate?

I think one of the important debates relates to the symbol grounding problem. How do you actually establish, if you know with a symbol based approach, how do you establish a relation of the symbol to the outside world. That is the symbol grounding problem. Maybe if you pursue a bottom-up approach as we are doing, trying to build cognition from scratch, then you do not run into the symbol grounding problem. But in this way it is more difficult to get the systems to do what they are supposed to do.

So I think there is this kind of trade-off that people often say: I want this system to do this and this and this which is often the case in cognitive systems research and then they basically program the systems. That is fine if it works for a particular application. The problem with this is that you do not learn very much about cognition in general when you do this. You may be able to develop useful applications which are what many people want, but I think at least one of our main goals is to really understand the basic principles underlying cognition. It may be the case that it will be a longer and harder journey until we get there but I think it is the more interesting one. At least from our perspective.

Thank you!

Text only
An euCognition Network Action
Department of Computer Science
University of Bath
Send us an email:
euCognition -
HomeFor PressFor KidsFor AcademicsAbout Us
Meet the Scientists
Explore the Debates
Study at University
Robot Gallery
Home > Meet the Scientists > Prof. Rolf Pfeifer