We present an experimentation system we develop at the University of Hamburg for the study of space and time and spatio-temporal representations. To motivate our approach, we first review principal approaches to the study of cognition. We then explain what we call the realator approach by reflecting on the notion of reality. Finally, we present the system and its components and conclude with a brief report of some intial results. 1 Approaches to the Study of Cognition In studying principles of cognitive function, a variety of models has been proposed and used. Figure 1 serves to illustrate these models. For the purpose of this discussion, we assume that we have a cognitive task in the "real world" (left top node RP in the figure), i.e. a problem that is to be solved. The solution of the problem (in the "real world" ) is depicted by the right top node RS in the figure. In performing a task (or solving a problem) natural cognitive beings traverse a path from RP to BARKOWSKY, T., BERENDT, B., EGNER, S., FREKSA, C., KRINK, T., RÖHRIG, R., WULF, A. 1994. The Realator How to Construct Reality -. In: Proc. ECAI-94 Workshop on Spatial and Temporal Reasoning. pp. 19-26 RS, i.e., they carry out actions in the "real world". We will call this type of cognition here "natural cognition". An example of natural cognition may be the task of opening a closed door; RP corresponds to the state in which the door is closed, RS corresponds to the state in which the door is open, in between some action occurred which caused the transition from RP to RS. Of course, this action may have happened in such a way that we do not want to view it as a cognitive action for example when the wind opened the door, or when seemingly random actions of an animal cause the door to open. In the first case, the physical process appears too straightforward to qualify for being considered cognitive, in the second case it may be so complicated that we do not recognize the structure; thus, it does not qualify for cognition either. In between these two extremes we have those actions which appear controlled by some instance 'that knows what it is doing'; these actions we are inclined to label 'cognitive actions'. Knowing is related to representation. In order to better understand knowledge, artificial intelligence makes use of formal (artificial) representations of reality: the real problem is represented by an abstract pendant (left bottom node AP), an action is performed in abstract space, an abstract solution (right bottom node AS) is obtained; this solution is translated into the corresponding state RS. Besides the direct link between RP and RS (corresponding to natural cognition) we can identify three other links: the "representation" link from RP to AP, the "artificial cognition" link from AP to AS, and the link from AS to RS we will call it "interpretation" link (cf. [Palmer 78]). The only actions which take place entirely in the abstract domain are the actions corresponding to the artificial cognition link. The representation and interpretation links connect the real and the abstract worlds and thus involve both domains. In artificial intelligence, cognition is studied by focusing on the artificial cognition link: assuming that the representation and interpretation tasks are trivial or easy, the cognition problem has been identified with the abstract cognition task. Neglecting the representation and interpretation links in essence identifies the natural cognition task with the artificial cognition task. There are two classes of approaches to tackling the artificial cognition problem: the first exclusively studies the formal aspects of the problem and develops solutions independently of solutions in the "real world". The "logic approach" belongs to this class. The second class studies the BARKOWSKY, T., BERENDT, B., EGNER, S., FREKSA, C., KRINK, T., RÖHRIG, R., WULF, A. 1994. The Realator How to Construct Reality -. In: Proc. ECAI-94 Workshop on Spatial and Temporal Reasoning. pp. 19-26 natural cognition processes and simulates them in the artificial domain. Cognitively oriented researchers using classical AI methods usually subscribe to this class of approaches. There is another view of cognition which considers the artificial cognition link as easy and uninteresting, once a good representation is available; the challenge of cognition here consists in ways of finding representations which will make the abstract cognition task easy or trivial. This approach is confronted with two rather different worlds: with the "real world" and its processes and with the abstract world and its operations. In effect, all four links have to be considered simultaneously; thus, there are only few constants to deal with. The difficulty of establishing correspondences between real situations and their abstract representations has led some computer scientists to radically reject representations [Brooks 91]; they advocate studying cognition directly within the real situations in which the tasks come about. We will call this approach the "situation approach". By avoiding abstraction, it makes all aspects of the real situation with full complexity available. This may be an advantage, as important influences of unknown origin may be detected; on the other hand, it may make it hard to isolate simple principles. BARKOWSKY, T., BERENDT, B., EGNER, S., FREKSA, C., KRINK, T., RÖHRIG, R., WULF, A. 1994. The Realator How to Construct Reality -. In: Proc. ECAI-94 Workshop on Spatial and Temporal Reasoning. pp. 19-26 2 The Realator Approach There is an alternative to both, the simulation approach and the situation approach to cognition: we can combine their advantages and study cognitive processes completely within an artificial reality. By creating artificial environments, we can control the contributing factors to our domain, as we define the domain as we do in formal systems. On the other hand, we avoid the representation and interpretation problems since we do not claim that we solve problems in the natural reality. In effect, we collapse artificial cognition with natural cognition by identifying the nodes AP with RP and AS with RS. So far, we differentiated between two 'realities', the natural and the artificial one. Of course, when we commonly refer to reality, we actually mean 'natural reality'. Thus, as we are in our further considerations not only concerned with the two realities already mentioned, but additionally with a third one, the one that occurs when natural and artificial cognition are collapsed into one, it is useful to determine, what the constituting components of reality are in general. Why are we sure to be concerned with 'reality', when we speak about our usual natural reality? In fact, we call 'real' what anyone of us would describe to perceive just the way we do, i.e. what can be seen, heard, smelled, tasted etc. by our senses and what leads to a description of perception which is compatible to that one of other humans. So we can say, that something is real, because it appears to us the way we are used to by means of our sensory properties. Or, with respect to the world, we can perceive everything our senses inform us about. Indeed, there are aspects of the world, we cannot detect with our senses directly, e.g. any energetic radiation besides the wavelength of visible light. In this case, we need a helping medium that transforms those aspects of the world not directly accessible to some domain examinable by our senses. In consequence, what we call 'reality' are those aspects of the world we can come to know about by perception. That means, our reality is made of the environment we live in and our senso-motoric capabilities with respect to this environment. Our 'real world' is this part of our environment, that is senso-motorically accessible for us. Departing from this notion of reality, we create a world for thought experiments and experimentation similar to that of Braitenberg's "vehicles" [Braitenberg 84]; but by synthesizing 'spaces', we emphasize environmental aspects or 'situations' that perceiving and acting entities BARKOWSKY, T., BERENDT, B., EGNER, S., FREKSA, C., KRINK, T., RÖHRIG, R., WULF, A. 1994. The Realator How to Construct Reality -. In: Proc. ECAI-94 Workshop on Spatial and Temporal Reasoning. pp. 19-26 are confronted with. To contrast our approach with the simulation approach we have labeled our system the "realator". That will say, that with our experimentation system we construct a reality in the sense explored above: we generate an (artificial) environment and an entity interacting senso-motorically with this environment. The fact, that the environment is not 'natural' but artificial, is of no importance when concerned with reality construction. Indeed, as only the sensorial access that a being has to its environment is of importance for generating its reality, it is of no use asking what this environment means to an external observer (i.e., if someone else considers the environment to be natural or artificial). By studying an artificial entity in an (from our viewpoint) synthetic environment, we expect to learn much about the way we deal with our own natural environment. In particular, we can examine the interdependencies between the perceivable reality and the concepts used for cognitively motivated action within this reality. By varying parameters in our artificial reality, we can learn about critical aspects for modeling cognitive systems in natural environments. But by studying artificial worlds, we do not claim to solve the difficult representation and interpretation problems along with the reasoning problem.