Sociable robots are gaining popularity among robotics researchers and cognitive scientists. These robots generally show abilities that include expressive power (face, voice, ...), locating, paying attention to and addressing people, etc. Such abilities fall into what is known as social intelligence in humans. The reproduction of social intelligence, as opposed to other types of human abilities, may lead to fragile performance, in the sense of having rather different performances between tested cases and future (untested) cases and situations. This limitation stems from the fact that our social abilities are mainly unconscious to us. This is in contrast with other human abilities that we carry out using conscious effort, and for which we can easily conceive algorithms and representations. The fragile performance mentioned above is nothing but overfitting. Thus, we propose to approach the problem using strategies followed in Machine Learning for avoiding overfitting. Specifically, complexity penalization and incremental design are translated to the broader ambit of robot design and development. The robot CASIMIRO is currently being developed following that approach.