689 International Journal on Advances in Intelligent Systems, vol 7 no 3&4, year 2014, http://www.iariajournals.org/intelligent_systems/ 2014, © Copyright by authors, Published under agreement with IARIA - www.iaria.org Modelling Spatial Understanding: Using Knowledge Representation to Enable Spatial Awareness and Symbol Grounding in a Robotics Platform Martin Lochner, Charlotte Sennersten, Ahsan Morshed, and Craig Lindley CSIRO Computational Informatics (CCI) Autonomous Systems (AS) Commonwealth Scientific and Industrial Research Organization (CSIRO) Hobart, Tasmania, Australia Contact: martin.lochner@csiro.au, charlotte.sennersten@csiro.au, ahsan.morshed@csiro.au, craig.lindley@csiro.au Abstract—Robotics in the 21st century will progress from scripted interactions with the physical world, where human programming input is the bottleneck in the robot’s ability to sense, think and act, to a point where the robotic system is able to autonomously generate adaptive representations of its surroundings, and further, to implement decisions regarding this environment. A key factor in this development will be the ability of the robotic platform to understand its physical space. In this paper, we describe a rationale and framework for developing spatial understanding in a robotics platform, using knowledge representation in the form of a hybrid spatial- ontological model of the physical world. Further, we describe the proposed CogOnto (cognitive ontology) model, which enables symbol grounding for a cognitive computing system, using sensor data gathered from diverse and heterogeneous sources, associated with humanly crafted symbolic descriptors. While such a system may be implemented with classical ontologies, we discuss the advantages of non-hierarchical modes of knowledge representation, including a conceptual link between information processing ontologies and contemporary cognitive models. Keywords-Human Robot Interaction; Artificial Intelligence; Autonomous Navigation; Knowledge Representation; Symbol Grounding; Spatial Ontology. I. INTRODUCTION The process of transitioning away from hard-coded robotics applications, which carry out highly pre-determined actions such as the traditional manufacturing robot, is already well underway. This paper follows our previous work [1] in which we describe a methodology for using ontological data representation to encode 3D spatial information in robotics applications. With notions such as cloud robotics [2] entering the zeitgeist, and highly publicized events such as the Defense Advanced Research Projects Agency (DARPA) Robotics Challenge (Dec. 19-21, 2013, Miami FL) bringing public attention to these advances, it is foreseeable that robots will be entering the mainstream realm of human activity – more than in fringe applications (robotic vacuum cleaner; children’s toys), but in key areas such as caring for the aged [3], operating vehicles [4], disaster management [5], and undertaking autonomous scientific investigation [6]. The hurdles that must be overcome in reaching these goals, however, are neither few nor small. This can be plainly seen, for example in the aforementioned 2013 Robotics Challenge, in which simple spatial tasks that are routine for a human being (open a door, climb a ladder) are still critically difficult for even the most advanced and highly funded robotics projects. While the state-of-the-art is impressive, it is evident that physical robotics hardware is far in advance of the control systems that are in place to guide the robot. The challenge is, thus, to develop systems whereby a robot can perceive a physical space and understand its position in that space, the components that exist within the space, and how it can or should interact with these components in order to achieve implicit or explicit goals. This is furthermore impacted by the requirement that robotic systems be able to operate in outdoor environments where distributed connections may not be available; however, describing the development of long-range data networks for robotic communication is beyond the scope of this paper. While there are a number of ways that the problem of providing a robot with a spatial understanding can be approached (e.g., neuro-fuzzy reasoning [7], dynamic spatial relations via natural language [8]) it is our proposition that leveraging the current advancements in knowledge representation via ontologies [9][10], in combination with an understanding of human spatial- cognitive processing [11][12], and enabled by real-time scene modeling [13] will provide a powerful and accessible methodology for enabling spatial understanding and interaction in a mobile robotics platform. As argued by Sennersten et al. [14], the advantage of using cloud-based repositories of perceptual data annotated with ontology and metadata information is to take advantage of humanly- tagged examples of sense data (e.g., images) to overcome the symbol grounding problem. Symbol grounding refers to the need for symbolic structures to have valid associations with the things in the world that they refer to. Achieving symbol grounding is an ongoing challenge for robotics and other intelligent systems [15]. Using cloud-based annotations attached to sensory exemplars takes advantage of the human ability to ground symbols, obviating the need