Towards Persistent Storage and Retrieval of Domain Models using Graph Database Technology  Nico Hochgeschwender, 1 , 2   Holger Voos, 2   and Gerhard K. Kraetzschmar 1  Robots are expected to perform a wide range of challeng- ing tasks in dynamic environments. To do so robots need to extract knowledge about the world from the data perceived through the sensors of the robot. However, as Crowley   et al.   [1] already pointed out the huge variations in operating conditions such as environmental changes (e.g. illumination and occlusion) and failures (e.g. sensor failures) makes the design, development and deployment of robot perception ar- chitectures a challenging and knowledge-intensive exercise. To easily configure, modify and validate robot perception architectures we proposed in our previous work the domain- specific   language   RPSL   (Robot   Perception   Specification Language) [2]. The RPSL is a Ruby-based internal DSL which enables to specify individual configurations (called  perception graphs ) of RPAs meeting the requirements of different context conditions. To this end RPSL enables to specify two crucial elements of perception systems, namely perception graphs and data types. In RPSL a perception graph is a composition of components distinguished in sensor and processing components in the form of a directed acyclic graph (DAG). As the RPSL is an internal DSL the resulting  domain models   of concrete perception graphs are Ruby code itself. Up to now we stored these domain models in a self- made repository implemented with standard Ruby collections such as hashes, sets, and ranges. We used the repository mainly to facilitate the selection of perception graphs during runtime [3]. However, as the repository is steadily growing e.g., new perception graphs are added or existing perception graphs are modified a more structured approach to store and retrieve domain models is required. A common approach to organise data in a structured way is to employ some sort of database management system (DBMS). In the context of this work we decided to employ a graph database (GD) 1   to store and retrieve RPSL domain models. A GD stores information in the form of a graph where nodes represent entitites and edges among nodes represent relations. As RPSL domain models are already structured as a graph the decision to choose a GD as DBMS is feasible. Beyond that, a GD enables to express so called   semantic queries   in order to  *Nico Hochgeschwender is recipient of a PhD scholarship from the Graduate Institute of the Bonn-Rhein-Sieg University, which he gratefully acknowledges.  1 Nico   Hochgeschwender,   Artem   Vinokurov,   and   Gerhard K.   Kraetzschmar   are   with   the   Department   of   Computer Science,   Bonn-Rhein-Sieg   University,   Sankt   Augustin,   Germany.  nico.hochgeschwender@h-brs.de .  2 Nico Hochgeschwender and Holger Voos are with the University of Luxembourg, Luxembourg.  1 We make use of the   http://neo4j.com/   graph database.  retrieve and infer information based on the relations among the domain models. By doing so we can link the   domain models   as exemplified in the following simplified RPSL repository. Let us assume we have two perception graphs, each composed of two components (see Fig.. 1). In both perception graphs the first component produces some output conforming to a type   T   . Here,   T   is a separated RPSL domain model dealing with data type information. In our repository we link both domain models as seen in the Figure. This  Fig. 1.   Simplified example of a perception graph repository  enables use to perform sophisticated queries incorporating both domain models such as retrieve all the components producing type   T   , or get all the perception graphs where at least one component produces type   T   . As we use the  neo4j   GD we can utilize the   Cypher DSL   a dedicated SQL-like query language for GDs. For instance, to retrieve all the components producing type   T   the following Cypher statement is sufficient:  match (n:Component), (m:Type) where (n)-[:of_type]->(m) return n;  Up to now this language and the GD approach in general is promising. Roughly speaking, it enables to perform two gen- eral types of queries, namely meta-level queries to retrieve general information (e.g. number of components, graphs etc.) about the repository and semantic queries incorporating several domain models. R EFERENCES [1]   Crowley, James L. and Hall, Daniela and Emonet, Remi.   Autonomic Computer Vision Systems . The 5th International Conference on Com- puter Vision Systems. 2007. [2]   Hochgeschwender, Nico and Schneider, Sven and Voos, Holger and Kraetzschmar, Gerhard.   Declarative Specification of Robot Percep- tion Architectures . Simulation, Modeling, and Programming for Au- tonomous Robots. Volume 8810 of the series Lecture Notes in Com- puter Science. 2014. [3]   Hochgeschwender, Nico and Olivares-Mendez, Miguel A. and Voos, Holger and Kraetzschmar, Gerhard.   Context-based Selection and Ex- ecution of Robot Perception Graphs . IEEE International Conference on Emerging Technologies and Factory Automation (ETFA). 2015.  arXiv:1607.04138v1 [cs.RO] 19 Feb 2016