Learning intermediate object affordances: Towards the development of a tool concept

Inspired by the extraordinary ability of young infants to learn how to grasp and manipulate objects, many works in robotics have proposed developmental approaches to allow robots to learn the effects of their own motor actions on objects, i.e., the objects affordances. While holding an object, infants also promote its contact with other objects, resulting in object-object interactions that may afford effects not possible otherwise. Depending on the characteristics of both the held object (intermediate) and the acted object (primary), systematic outcomes may occur, leading to the emergence of a primitive concept of tool. In this paper we describe experiments with a humanoid robot exploring object-object interactions in a playground scenario and learning a probabilistic causal model of the effects of actions as functions of the characteristics of both objects. The model directly links the objects' 2D shape visual cues to the effects of actions. Because no object recognition skills are required, generalization to novel objects is possible by exploiting the correlations between the shape descriptors. We show experiments where an affordance model is learned in a simulated environment, and is then used on the real robotic platform, showing generalization abilities in effect prediction. We argue that, despite the fact that during exploration no concept of tool is given to the system, this very concept may emerge from the knowledge that intermediate objects lead to significant effects when acting on other objects.