An Embodied Model for Handwritten Digits Recognition in a Cognitive Robot

This paper presents an embodied model for recognition of handwritten digits in a cognitive developmental robot scenario. Inspired by neuro-psychological data, the model integrates three modules: a stacked auto-encoder network to process the visual information, a feedforward neural controller for the fingers, and a generalized regression network that associates number digits to finger configurations. Results from developmental learning experiments show an improvement in the digits' recognition rate thanks to the inclusion of the robot fingers in the training especially in its early stages (epochs) or with a low number of examples. This behaviour can be linked to that observed in psychological studies with children, who seem to benefit of finger counting only in the initial stage of mathematical learning. These results suggest the potential of the embodied approach to favour the creation of a psychologically plausible developmental model for mathematical cognition in robots and to support the creation of more complex models of human-like behaviours.