In this paper, we study how a biologically-plausible cerebellum architecture can store and retrieve different robotic-arm internal models (in synaptic connections between granular layer and Purkinje cells) at the granule layer (dynamic modifications of a base robot-arm-plant model), and how the model microstructure and input signal representations can efficiently infer models in a robot control scenario during object manipulation. More specifically, we have evaluated the contribution of the granular layer to the ability of the cerebellum to generate corrective actions. To achieve this we have embedded a spiking cerebellar model into an analog control loop whose output commands a simulated robot arm. The performance results obtained by using a cerebellum which includes granular layer are compared to those using a cerebellum without this layer. The results show that this layer effectively contributes to the generation of accurate cerebellar corrections. This work represents a well defined case of study in the field of neurobotics, in which biologically plausible neural systems and robots are used to study the functionality of biological systems.