Exploring uncertainty and movement in categorical perception using robots

Abstract

Cognitive agents are able to perform categorical perception through physical interaction (active categorical perception; ACP), or passively at a distance (distal categorical perception; DCP). It is possible that the former scaffolds the learning of the latter. However, it is unclear whether ACP indeed scaffolds DCP in humans and animals, nor how a robot could be trained to likewise learn DCP from ACP. Here we demonstrate a method for doing so which involves uncertainty: robots are trained to perform ACP when uncertain and DCP when certain. We found evidence in these trials that suggests such scaffolding may be occurring: Early during training, robots moved objects to reduce uncertainty as to their class (ACP), but later in training, robots exhibited less action and less class uncertainty (DCP). Furthermore, we demonstrate that robots trained in such a manner are more competent at categorizing novel objects than robots trained to categorize in other ways.