Abstract
This paper presents a biologically-inspired model employing bottom-up visual attention for robot task learning. Although bottom-up attention enables robots to detect likely important information, discontinuity of the attention as well as its instability causes a challenge in being applied to action learning. The proposed model overcomes the problem by examining spatial and temporal continuity in low-level features for attended locations. Retina filtering and stochastic attention selection, which are integrated with saliency-based visual atteneion, facilitate the process by stabilizing the model's attention while keeping its receptiveness to a new stimulus. An experiment shows that the model can extract key actions from task demonstration.
