Complex affordance learning based on basic affordances

In this paper, we study how complex object affordances can be efficiently learned and how previously learned structures can be used for this purpose. We discuss that besides standard visual features, using previously learned basic affordances in predicting complex affordances would speed up this complex learning task. In order to prove our hypothesis, we compared two different types of complex affordance predictors: The predictors that are based on shape features and the ones that use basic affordances. The results obtained from a synthetic (object, action) interaction database showed that basic-affordance based predictors can generalize over novel objects even with small training sets. This result shows that although the basic affordances are related to basic simpler actions, as they encode object-robot-environment dynamics, they can speed up learning of complex actions.