Reinforcement Learning of Predictive Features

Recently, the aspect of visual perception has been explored in the
 context of Gibsons concept of affordances [4] in various ways. In
 extension to functional views on visual feature representations,
 we focus on the importance of learning in perceptual cueing for an
 agents anticipation of opportunities for interaction. In this context
 we demonstrate the learning of causal relations between visual cues
 and associated anticipated interactions by reinforcement learning
 of predictive perceptual states. We pursue a recently presented framework
 for cueing and recognition of affordance-based visual entities that
 obviously plays an important role in robot control architectures,
 in analogy to human perception. We verify the concept within a real
 world robot scenario by learning predictive visual cues via reinforcement
 signals, proving that features were selected for their relevance
 in predicting opportunities for interaction.