In this paper, an architecture for an emotion-based agent and its application to a real robot, situated in a semistructured environment, is presented. This architecture corresponds to an improved version of the one proposed by Maqas et. al.. Both were developed based on a particular interpretation of the neuro-physiological findings of Damisio and LeDoux, namely the concepts of stimuli parallel processing, by LeDoux, and somatic marking, by Damisio. Given several applications of this architecture with virtual agents and in simulated environments, the goal here was to study and evaluate the utility and efficiency of this emotion-based agent architecture with real robots and real environments.