Convolutional Neural Networks have brought in the last years a significant advance in the perceptual visual abilities of intelligent autonomous systems like robots. Still, these algorithms need to 'overspecialize' in order to achieve performances robust enough to be employed in realistic settings. While this is a suitable strategy for methods design to perform a single visual task, it is clearly suboptimal in the case of a robot system, where multiple and varying visual tasks are requested in order to act and interact intelligently with the environment. The goal of this thesis is to study the problem of learning without forgetting in the robot vision context, developing a benchmark evaluation protocol to assess the advantages and weaknesses of possible approaches, and evaluate the current state of the art in computer vision in this new, challenging scenario..