Bootstraping manipulation skills to learn affordances in open-ended environments*

A cause-effect relation is called causality [1]. The causality relations with objects in an environment are known as affordances, defined as the acquired cause-effect relation of applying a behaviour/action on an entity/object to obtain an effect [2]. Affordance knowledge is part of the expertise needed by robots to properly perform tasks in an environment. This expertise is threefold: (i) knowledge related to the robot itself, associated to the concept of self and some basic motion capabilities, i.e., its kinematic model; (ii) knowledge related to interpreting its environment, e.g., segmentation, identification and tracking of objects; (iii) capabilities to interact with the environment, composed by (a) affordance knowledge, to infer the right actions to perform given an environmental setup; and (b) manipulation skills, motion primitives as grasp and push, to execute those actions. To the best of the authors’ knowledge, in the literature related to affordances the experimental setups are closed, i.e. composed by a set of preselected objects. In these environments robots are usually provided with both the information about themselves and the information to interpret the environment, as acquiring the corresponding skills is (an extended review of this literature is available in [3]). Regarding the interaction with the environment, robots are also provided with a set of predefined motion primitives adapted to produce effects on the previously selected objects. For example, a robot could be asked to clutter a table with some objects on it. Therefore, it should be provided with a set of primitives to be able to properly arrange them. However, in daily situations, environments are usually open-ended, i.e., only partially known, with incomplete information. In these environments the definition of general purpose primitives must be complex, due to the unforeseen situations a robot can face. On the previous cluttering scenario, if a soft object, as a cuddly toy, is put on the table the robot would not be able to adapt to the new environment, i.e., it would not be able to produce a desired effect (e.g., push to the left) on this object. Developmental Robotics aims at making robots acquire their own knowledge through their