Declarative Specification of Robot Perception Architectures Nico Hochgeschwender 1, 2 , Sven Schneider 1 , Holger Voos 2 , and Gerhard K. Kraetzschmar 1 1 Bonn-Rhein-Sieg University, Computer Science, Sankt Augustin, Germany nico.hochgeschwender@h-brs.de 2 University of Luxembourg, SnT Automation Research Group, Luxembourg Abstract. Service robots become increasingly capable and deliver a broader spectrum of services which all require a wide range of perceptual capabilities. These capabilities must cope with dynamically changing re- quirements which make the design and implementation of a robot percep- tion architecture a complex and tedious exercise which is prone to error. We suggest to specify the integral parts of robot perception architec- tures using explicit models, which allows to easily configure, modify, and validate them. The paper presents the domain-specific language RPSL, some examples of its application, the current state of implementation and some validation experiments. 1 Introduction Service robots operating in industrial or domestic environments are expected to perform a wide variety of tasks in different places with often widely differing environmental conditions. This poses many challenges for the perception-related parts of the control software (here referred to as robot perception archi- tecture (RPA)), which includes recognizing and tracking manipulable and non- manipulable objects, furniture, people, faces, and recognizing gestures, emotions, sounds, and speech. Designing a single set of perception components that performs all these per- ceptual tasks simultaneously, robustly, and efficiently would require enormous effort and would result in unmanageable complexity. To meet the challenges of service robotics we need concepts, methods, and tools for designing and devel- oping RPAs in a very flexible manner. Ultimately the robot should be able to adjust to the wide range of situations autonomously (e.g. by dynamically select- ing a set of perceptual components into an RPA configuration for a particular task). Fig. 1 illustrates the concept, where the components shown in red com- pose the RPA configuration active when the pose of a person is required. To do so, explicit knowledge representation about its available perception capabili- ties/functionalities (as depicted in Fig. 1) are required. However, many RPA design decisions remain nowadays implicit. These deci- sion concern the robot platform, robot’s tasks, and the environment in which the