Toward Autonomous Learning of an Ontology of Tool Affordances by a Robot

complex environments, however, it is impossible to know in advance the exact nature and number of possible environmental outcomes that the robot can induce through its behaviors. In addition, the changes that the robot can detect are often high-dimensional, making it difcult to use standard machine learning algorithms. This work addresses this problem by proposing a framework in which the robot learns a taxonomy for the types of perceivable changes produced by its own behaviors. The proposed method also allows the robot to incrementally update the taxonomy and to conceptualize new types of observed outcomes. In addition, the robot solves a hierarchical classication task by learning a model that predicts the future outcome of its behaviors in relation to the learned taxonomy. Thus, the robot builds an affordance ontology consisting of an outcome class taxonomy and a predictive model grounded in the robot’s perceptual and behavioral repertoire.