Learning Object Affordances from Sensory-Motor Interaction via Bayesian Networks with Auto-Encoder Features

Abstract

In this paper, we study learning relationships between objects, actions, and effects. "Affordance" is an ecological psychology concept, which addresses how humans learn these relationships and which is also studied in cognitive robotics to transfer the same ability to robots. Our model is built on top of two existing models in this field and uses their strengths to introduce a novel system, where an anthropomorphic robot observes its environment and changes in that environment after executing pre-learned actions. Robot transfers these observations to object and effect properties in the same space and object affordances are learned using Bayesian Networks. The dimensions of features are decreased through autoencoders to achieve a compact network. Usage of a probabilistic model helps our system to deal with missing information or to make predictions for object properties and actions along with effect properties. We illustrate the advantages of our model by comparing it with the two aforementioned models.