There are two main problems of the robot architecture. Firstly, a suitable approach to the realization of behaviors has to be chosen and its appropriateness and feasibility for the given scenario has to be shown. In this paper 'Neural Fuzzy Controller In Behaviour-Oriented Architectures', Steffen Förster describes an improvement of the architecture for reactive systems proposed by Luc Steels. Förster shows that the advantages of two independent technologies, artificial neural networks and the theory of fuzzy systems, can be exploited to enhance the ease of definition and the adaptivity of Steels-style reactive systems significantly.
The second problem our project is confronted with concerns with the interface between the reactive behavior system and a symbolic cognitive system on top of it, that contains the natural language system and the planning component. The philosophy of behavior-based systems defy straightforward combination with goal-oriented planning systems. Jan-Torsten Milde discusses in his paper 'An Architecture Proposal for the Integration of a Reactive and Cognitive System' the interaction of behavior and planning systems that is made possible by an hierarchical organization of behavior together with explicitly represented information on activation and selection state of each behavior.
Due to the dominant role of the processing of sensory data from different sources in our approach to an instructable though during execution still reactive system, we decided to develop a simulation system for the complete scenario as a testbed for our architecture. The simulation system allows to equip the robot very easily with numerous sensors of different types and to investigate their interplay during task execution. Gil Müller describes the architecture of a simulation system that contains as its core the so-called I-Space, a communication tool for multi-agent systems.
The process of design has been made easier without the loss of functionality. The system is capable of learning with already known algorithms (e.g. backpropagation, reinforcement learning).
The hybrid system is used for the control of a simulated robot arm, that has to construct a toy plane. Instructions are given to the robot arm in natural language. Besides the capacity of natural language understanding, the arm is equipped with a set of sensor for the perception of visual and tactile information.
I-Space is responsible for the transportation of messages in a manner that is indepentent from the actual physical network. To use its service, applications (we will call them objects) have to connect to it. Then they can send and receive messages. The transport of those messages depends on three parameters: receiver, message type and time. The receiver of such a message could be anyone on a range of all connected objects to one particular object. This means that each specified receiver will get a copy of the message. This could result in a great amount of messages sent. Filtering is a means to restrict that. The system supports simple filters through message types. Thus, only those messages will be delivered to a particular object that match the message types the object has specified. In real-time systems messages have to arrive within a fixed amount of time. I-Space suuports that through time-based scheduling. The object can control that mechanism by specifying the time of delivery. Messages are delivered on demand of the connected objects (i.e. by polling) or by signaling the objects (by interrupt).