Development of Categorisation Abilities in Evolving Embodied Agents: A Study of Internal Representations with External Social Inputs

The present paper investigates how embodied and situated agents perform tasks that require skills of categorization. The agents are artificially created by an adaptation process. The task is to categorize different shapes of objects using sensory-motor and linguistic input. Results show that the autonomous agents are able to solve the categorization task by integrating the sensory-motor experienced states and employing “linguistic” input from the environment. This shows that autonomous agents are able to develop some "emerging" abilities by exploiting the information present in the environment in order to recognize and discriminate objects. Autonomous agents also exhibit a "social" behavior, because they are able to categorize the objects in the environment, even when external inputs are unavailable. The purpose of this work is to prove the theoretical hypothesis that the “social” information (external labels), deriving from another agent or from the trainer, facilitates individual capacity to categorize, by the creation of internal representations.