Bootstrapping paired-object affordance learning with learned single-affordance features

The aim of this paper is to propose a system where complex affordance learning is bootstrapped through using pre-learned basic-affordances as additional inputs of the complex affordance predictors or as cues in selecting the next objects to explore during learning. In the first stage, the robot learns affordances in the form of developing classifiers that predict effect categories given object features for different discrete actions applicable to single objects. These predictions are later added to robot´s feature set as higher-level affordance features. In the second stage, the robot learns more complex multi-object affordances using object and affordance features. We first applied our idea in an artificial interaction database which includes discrete actions, several manually coded object categories, and actions effects. Finally, we validated our bootstrapping approach in a real robot with poke and stack actions. We expected to obtain higher performance with affordance-features especially in small training datasets as the object-robot-environment dynamics should have already been partially learned and encoded in affordances. The experiment results showed that complex affordance learning significantly speeds up with predictors that are bootstrapped with affordance-features compared to predictors that use low-level features such as shape descriptors. We also showed that by actively selecting the next objects and by increasing the diversity of the training set using a distance measure based on learned single-object affordances, the effect of bootstrapping can be further increased.