Cerebellar Spiking Engine: Towards Objet Model Abstraction in Manipulation N. R. Luque, J. A. Garrido, R. R. Carrillo, E. Ros Abstract-This paper presents how a plausible cerebellum- like architecture can abstract corrective models in the framework of a robot control task when manipulating objects that significantly affect the dynamics of the system. The presented scheme is adequate to control non-stiff-joint robots with low-power actuators which involve controlling systems with high inertial components. We evaluate the way in which the cerebellum stores a model in the granule layer, how its microstructure can efficiently abstract models and deliver accurate corrective torques for increasing precision during object manipulation. Particularly we study how input sensory- motor representations can enhance model abstraction capabilities during accurate movements, making use of explicit (model-related input labels) and implicit model representations (sensory signals). Finally we focus on how our cerebellum model (using a temporal correlation kernel) properly deals with transmission delays in sensory-motor pathways. Keywords-Spiking Neuron, Cerebellum, Adaptive, Simulation, Learning, Robot, Biological Control Systems I. I NTRODUCTION C ontrolling fast non-stiff-joint robots accurately with low power actuators is a difficult task which involves high inertia. Biological systems are in fact non-stiff-joint "plants" driven with relatively low-power actuators. These biological systems have developed smart "model abstraction" capabilities through evolution. In this way, the control commands of biological systems are generated taking into account the "plant model" (for instance arm + object). In the amework of accurate control with a large number of degrees of eedom (DOFs), extracting efficiently models om explorative manipulation, storing them without interference with other previously acquired ones and retrieving these models accurately for each case are Manuscript received February 5, 20 I O. This work has been supported by the EU project SENSOPAC (lST-028056) and the national grant DINAM- VISION (DPI2007-61683). N. R Luque and J. A. Garrido are PhD students at the University of Granada (Periodista Miguel Saucedo sIn Granada (Spain» in the Department of Computer Architecture and Technology. (E-mail: nlugue@atc.ugr.es and jgarrido@atc.ugr.es ). R. R Carrillo is a PhD researcher of the University of Almeria (Ctra. Sacramento sIn La Canada de San Urbano Almeria (Spain» in the Department of Computer Architecture and Electronics. (E-mail: rcarrillo@atc.ugr.es ). E. Ros is professor at the Univerity of Granada (Periodista Miguel Saucedo sIn Granada (Spain» in the Department of Computer Architecture and Technology. (E-mail: eros@atc.ugr.es). 978-1-4244-8126-2/101$26.00 ©2010 IEEE capabilities in which biological systems are still beyond current cutting edge control technology. With regards to machine learning approaches, specific models are being developed to address these hard tasks [1-4]. If the control scheme is based on accurate kinematic and dynamic models, and the dynamics of the plant changes (manipulating tools and objects will modi the base model), this will lead to significant distortions along the desired trajectory, affecting the final achieved accuracy [5- 7]. Therefore, these systems are required to be adaptive to tune the corrective models for specific object or tool manipulation [8]. The cerebellum seems to play a crucial role in this model abstraction task [9]. But how this is supported by actual network topologies, cell models and adaptation properties is an open issue. We have addressed the study of how this model abstraction task can be achieved with a cerebellum- like architecture based on spiking neurons. Lately, last generation simulation tools [10-14] allow the definition and simulation of nervous centers of certain complexity in the amework of biologically-relevant tasks. This allows addressing studies in which nction and structure of nervous centers are conjointly evaluated to better understand how the system operation is based on cell and network properties. In previous works we evaluated a cerebellar model in reaching point task [15] and a simple smooth pursuit task [16]. In this paper we evaluate how an input configuration encoding inherent propioception signals along the trajectory and also inputs om other sensory-systems (such as vision) are conjointly used in an object manipulation task. We also study how an adaptive cerebellum-like module using a basic temporal-correlation kernel (including long- term depression (LTD) and long-term potentiation (LTP) at parallel fiber-Purkinje cells synapses) can build corrective models to compensate deviations in the robot trajectory when the dynamics of the controlled plant is altered and also can deal properly with transmission delays in sensory- motor pathways.