Universität Bielefeld - Sonderforschungsbereich 360

Architectures in Cognitive Robotics

Bernd Hildebrandt, Reinhard Moratz and Christian Scheering (Eds.)

Introduction

Relevant aspects of the SFB 360 are concerned with architectures of artificial systems, especially with cognitive robots. At the 2. Workshop Kognitive Robotik of the "22. Jahrestagung Künstliche Intelligeny" (KI-98) held at the Universität Bremen (15. - 17. Sept. 1998) members of the SFB 360 presented their work and had some fertile exchanges of views with colleagues in related projetcs. The report presents a collection of papers on architectures in Cogntive Robotics that are based on lectures of the workshop. With this we intend to provide a varied basis for further discussion.

The contributions cover a wide range

All of them deal with several sources of information, perception as well as general world knowledge. Visual perception is without doubt the most prominent source of information. Often, serveral cameras acquire images of the same scene from different angels. Scheering et al. use sensor fusion to achiveve robutsness by taking advantage of redundancy, while the main interest of Becker et al. for using two cameras is the modelling of human visual perception. Beside vision other sources of sensor data (Scheering et al.), respectively of subsymbolic information (Goecke & Milde), are integrated.

The prominence of visual information processing is motivated by the necessity to recognize objects. For mobile robots objects are relevant as obstacles, i.e. as potential sources of collision, and as landmarks, i.e. as a source for positional information. Robots for manipulating objects need visual information to classify objects, i.e. to activate knowledge about objects, and to locate the three dimensional position of objects. Becker et al. supplement object recognition using gesture recognition. Here, predefined gestures are a means of instruction as joy sticks are used by Lankenau and Röfer.

Speech ist the most convenient means for humans to instruct other humans and, consequently, and artificial system. Verbal instructions are used for mobile robots (Moratz & Hildebrandt) as well as for assembly robots (Scheering et al.). The intention is that the user should be able to guide and assist the robot in a way they are used to. Becker et al. characterize such a robot system as semi-autonomous, i.e. "the robot must dispose of a repertoire of skills that are carried out autonomously, but the actual control of overall behaviour must be left to a human operator" (p. 4). The consideration of the communication between a robot and its instructor as the integral part of human-robot interaction leads to a conceptional extension of the classical viewof Cognitive Robotics. Since operational systems are not able to have sensible dialogues with humans in natural language Cogntive Robotics has mostly been concerned with robots acting autonomously.

The architectures chosen for these different kinds of robot systems depend on the intended performance and the hardware that is used. In addition to perceptual sources the architectures have to enable integration of objects knowledge for object manipulation, linguistic knowledge for verbal instructions, and spatial knowledge for navigation. On top of these the Bremen Autonomous Wheelchair requires a safe system architecture.

All robot systems that are presented in the papers have a modular architecture. An underlying hierarchy is motivated by different levels of information processing, e.g. levels of perceptual data processing, and (structured) levels of cognitive, respectively conceptual, representations and inference. Although modules are often appointed as agents (Becker et al.; Scheering et al.; Goecke & Milde), relevant differences can be found in the papers regarding the way and the complexity in which the communication between agents is implemented.

Anke Weinberger, 1999-03-16