Development of robot self-identification based on visuomotor prediction

We propose a developmental method that enables a robot to identify visual locations associated with its own body from a cluttered visual image based on the concept of visuomotor predictors. A set of statistical predictors are trained by linear regression to predict the visual features at each visual location from proprioceptive input. By measuring each predictor´s predictability using the R² statistics, the algorithm can determine which visual locations correspond to the robot´s body parts. Visual features are extracted using biologically plausible visual motion processing models. We demonstrate that while both orientation selective and motion selective visual features can be used for self-identification, motion selective features are more robust to changes in appearance.