Artificial Spatial Cognition for Robotics and Mobile Systems: Brief Survey and Current Open Challenges Paloma de la Puente and M. Guadalupe Sánchez-Escribano Universidad Politécnica de Madrid (UPM) Madrid, Spain Abstract—Remarkable and impressive advancements in the modeling general intelligent agents and their main areas of perception, mapping and navigation of artificial mobile commitments support the ambitious requirements of high level systems have been witnessed in the last decades. However, it is behavior in arbitrary situations for robotics [10]. A more recent clear that important limitations remain regarding the spatial model of spatial knowledge, the Spatial/Visual System (SVS) cognition capabilities of existing available implementations and [11] designed as an extension of the Soar cognitive the current practical functionality of high level cognitive models architecture, proposed a different multiplicity of [1, 2]. For enhanced robustness and flexibility in different kinds representations, i.e. symbolic, quantitative spatial and visual of real world scenarios, a deeper understanding of the depictive. The spatial scene is a hierarchy tree of entities and environment, the system, and their interactions -in general terms- their constitutive parts, with intermediate nodes defining the is desired. This long abstract aims at outlining connections transformation relations between parts and objects. Other between recent contributions in the above mentioned areas and research in cognitive architectures and biological systems. We try works in robotics employ similar internal representation ideas to summarize, integrate and update previous reviews, [12-14], and other ones included the possibility to hypothesize highlighting the main open issues and aspects not yet unified or geometric environment structure in order to build consistent integrated in a common architectural framework. maps [15]. While a complete implementation of this approach for all kind of objects requires solving the corresponding Keywords—spatial cognition; surveys; perception; navigation segmentation and recognition problems in a domain independent manner (which is far beyond the state of the art), I. BRIEF SURVEY keeping the perceptual level representations within the architecture enhances functionality. A very active research A. Initial models for spatial knowledge representation and community address these difficult challenges. main missing elements The recognition process should not only use visual, spatial Focusing on the spatial knowledge representation and and motion data from the Perceptual LTM but also conceptual management, the first contributions inspired by the human context information [7, 16] and episodic memories of cognitive map combined metric local maps, as an Absolute remembered places [17], from Symbolic LTM. This should Space Representation (ASR), and topological graphs [3]. As a also apply to the navigation techniques for different situations related approach, the Spatial Semantic Hierarchy (SSH) [4] [18, 19]. The existence of motion models for the objects can was the first fundamental cognitive model for large-scale improve navigation in dynamic environments, which is one of space. It evolved into the Hybrid SSH [5], which also included the main problems in real world robotic applications [20, 21]. knowledge about small-scale space. This fundamental work A novel cognitive architecture specifically designed for was undoubtedly groundbreaking, but it did not go beyond spatial knowledge processing is the Casimir architecture [22], basic levels of information abstraction and conceptualization which presents rich modeling capabilities pursuing human-like [6]. Moreover, the well-motivated dependencies among behavior. Navigation, however, has not been addressed, and different types of knowledge (both declarative and procedural) this work has scarcely been discussed in the robotics domain. were not further considered for general problem solving [7]. The SSH model was considered suitable for the popular One of the latest spatial models is the NavModel [23], schema of a “three layer architecture”, without explicitly designed and implemented for the ACT-R architecture. Besides dealing with processes such as attention or forgetting considering multi-level representations, this model presents mechanisms. This lack of principled forgetting mechanisms has three navigation strategies with varying cognitive cost. The been identified by the Simultaneous Localization and Mapping first developed implementation assumes known topological (SLAM) robotics community as a key missing feature of most localization at room level, while a subsequent implementation existing mapping approaches [8, 9]. incorporates a mental rotation model. This work focuses on the cognitive load and does not deal with lower level issues. B. The role of cognitive architectures and their relation to To point out how topics are addressed by the respective other works in the robotics community communities, we compiled Table I as a comparison. The Cognitive architectures provide a solid approach for contrast regarding memory management and uncertainty seems This work was partially funded by the Spanish Ministry of Economics and Competitivity: DPI 2014-53525-C3-1-R, NAVEGASE. It also received funding from the RoboCity2030-III-CM project (Robótica aplicada a la mejora de la calidad de vida de los ciudadanos. fase III; S2013/MIT-2748), supported by Programas de Actividades I+D en la Comunidad de Madrid and co-founded by Structural Funds of the EU. Proceedings of EUCognition 2016 - "Cognitive Robot Architectures" - CEUR-WS 52 to be relevant. The lack of approaches combining both [2] T. Madl, K. Chen, D. Montaldi and R. Trappl. Computational cognitive allocentric and egocentric representations is also remarkable. models of spatial memory in navigation space: A review. Neural Networks, 2015. To conclude, Table II shows a summary of surveys. [3] W.K. Yeap. Towards a computational theory of cognitive maps. Journal of Artificial Intelligence, 1988. TABLE I. COMPARISON OF TOPICS ADDRESSED BY THE COGNITIVE [4] B. Kuipers. The spatial semantic hierarchy. Artificial Intelligence. 2000. ARCHITECTURES AND ROBOTICS COMMUNITIES [5] Kuipers, J. Modayil, P. Beeson, M.MacMahon and F. Savelli. Local Cognitive Architectures ← Topic → Perception, Robotics, metrical and global topological maps in the hybrid Spatial Semantic Community Vehicles Community Hierarchy. ICRA, 2004. ACT-R/S, CLARION Egocentric spatial models [24, 25] [6] A. Pronobis and P. Jensfelt. Large-scale semantic mapping and reasoning with heterogeneous modalities. ICRA, 2012. LIDA, SOAR-SVS Allocentric spatial models [9, 26] [7] S.D. Lathrop. Extending cognitive architectures with spatial and visual Casimir, LIDA, SOAR-SVS Object based/ semantic representations [6, 12-14] imagery mechanisms. PhD Thesis, 2008. SOAR-SVS Explicit motion models / dynamic [27, 28] [8] J.A. Fernandez-Madrigal and J.L. Blanco. Simultaneous localization and information about the environment mapping for mobile robots: iIntroduction and methods. IGI, 2012. Memory management, forgetting [9] C. Cadena et al. Past, present, and future of simultaneous localization All [19] mechanisms and mapping: towards the robust-perception age. T-RO, 2016. Most mapping and Extended LIDA [29] Uncertainty considerations navigation approaches [10] U. Kurup and C. Lebiere. What can cognitive architectures do for robotics? Biologically Inspired Cognitive Architectures, 2012. [11] S.D. Lathrop. Exploring the functional advantages of spatial and visual cognition from an architectural perspective. TopiCS 2011. TABLE II. SUMMARY OF SURVEYS [12] R.F. Salas-Moreno, R.A: Newcombe, H. Strasdat, P.H.J Kelly and A.J. Davison. SLAM++: Simultaneous localisation and mapping at the Level Topic References of objects. CVPR, 2013. Robotics and Cognitive Mapping [1] [13] S. Eslami and C. Williams. A generative model for parts-based object segmentation. Advances Neural Information Processing Systems, 2012. SLAM and Robust Perception [8, 9] [14] A. Uckermann, C. Eibrechter, R. Haschke and H. Ritter. Real time Computational cognitive models of spatial memory [2] hierarchical scene segmentation and classification. Humanoids, 2014. [15] P. de la Puente and D. Rodriguez-Losada. Feature based graph SLAM in Object recognition [30, 31] structured environments. Autonomous Robots, 2014. Cognitive Architectures for Robotics [10] [16] L. Kunze et al. Combining top-down spatial reasoning and bottom-up Spatial knowledge in brains [17] object class recognition for scene understanding. IROS, 2014. [17] M.B Moser and E.I. Moser. The brain's GPS. Scientific American, 2016. [18] G. Gunzelmann and D. Lyon (2007) Mechanisms for human spatial competence. Spatial Cognition V, LNAI-Springer, 2007. II. CURRENT OPEN CHALLENGES [19] F. Dayoub, G. Cielniak and T. Duckett. Eight weeks of episodic visual The big challenge is closing the gap between high level navigation inside a non-stationary environment using adaptive spherical models and actual implementations in artificial mobile systems. views. FSR, 2013. To reduce this existing gap, we identify three main goals: [20] N. Hawes et al. The STRANDS project: long-term autonomy in everyday environments. Robotics and Automation Magazine, in press.  Combination of allocentric and egocentric models [21] P. de la Puente et al. Experiences with RGB-D navigation in real home using different levels of features/objects + robotic trials. ARW, 2016. topology/semantics. [22] H. Schultheis and T. Barkowsky. Casimir: an architecture for mental spatial knowledge processing. TopiCS, 2011.  Acquisition and integration of motion models and [23] C. Zhao. Understanding human spatial navigation behaviors: A dynamic information for the elements/objects. cognitive modeling. PhD Thesis, 2016.  Integration of global mapping & loop closure [24] R. Drouilly, P. Rives and B. Morisset. Semantic representation for capabilities with extensive declarative knowledge navigation in large-scale environments. ICRA, 2015. about features relevance and forgetting [25] L.F. Posada, F. Hoffmann and T. Bertram. Visual semantic robot mechanisms with episodic memory. navigation in indoor environments. ISR, 2014. [26] A. Richardson and E. Olson. Iterative path optimization for practical robot planning. IROS, 2011. ACKNOWLEDGMENT [27] R. Ambrus, N. Bore, J. Folkesson and P. Jensfelt. Meta-rooms: building The authors want to thank the EUCog community for and maintaining long term spatial models in a dynamic world. IROS, fostering interdisciplinary research in Artificial Cognitive 2014. Systems and organizing inspiring meetings and events. [28] D. M. Rosen, J. Mason and J. J. Leonard. Towards lifelong feature- based mapping in semi-static environments. ICRA, 2016. [29] T. Madl, S. Franklin, K. Chen, D. Montaldi and R. Trappl. Towards real- world capable spatial memory in the LIDA cognitive architecture. REFERENCES BICA, 2016. [30] J. J. DiCarlo, D. Zoccolan and N. C. Rust. How does the brain solve [1] G. Eason M. Jefferies and W.K. Yeap. Robotics and cognitive visual object recognition? Neuron, 2012. approaches to spatial mapping. Springer, 2008. Proceedings of EUCognition 2016 - "Cognitive Robot Architectures" - CEUR-WS 53