{agoncalves, jabrantes, gsaponaro, ljamone, alex}@isr.ist.utl.ptAbstract —Inspired by the extraordinary ability of younginfants to learn how to grasp and manipulate objects, manyworks in robotics have proposed developmental approaches toallow robots to learn the effects of their own motor actions onobjects, i.e., the objects affordances. While holding an object,infants also promote its contact with other objects, resulting inobject–object interactions that may afford effects not possibleotherwise. Depending on the characteristics of both the heldobject (intermediate) and the acted object (primary), systematicoutcomes may occur, leading to the emergence of a primitiveconcept of tool. In this paper we describe experiments witha humanoid robot exploring object–object interactions in aplayground scenario and learning a probabilistic causal modelof the effects of actions as functions of the characteristics of bothobjects. The model directly links the objects' 2D shape visualcues to the effects of actions. Because no object recognitionskills are required, generalization to novel objects is possibleby exploiting the correlations between the shape descriptors.We show experiments where an affordance model is learned ina simulated environment, and is then used on the real roboticplatform, showing generalization abilities in effect prediction.We argue that, despite the fact that during exploration noconcept of tool is given to the system, this very concept mayemerge from the knowledge that intermediate objects lead tosignicant effects when acting on other objects.