複数センサデータの不変性に基づく身体の発見 Body Finding based on the Invariance in Multiple Sensory Data

Finding the body in uninterpreted sensory data is one of the fundamental competences to construct the body representation that influences on adaptability of the robot to the changes in the environment and the robot body itself. The invariance of sensation seems a promising key information to find the self body since the sensory data are considered to be consistent in self body observation. To discriminate its body from non-body, the robot should complementarily utilize the invariance in multiple sensory data since single sensory data involve noise or a certain ambiguity occurred in the observation process. In this paper, we propose a method to discriminate body from non-body based on a conjecture about the distribution of the variance of sensations in terms of each observing posture. It can be approximated by a mixture of two Gaussian distributions for observing the body and non-body, respectively. After estimating the distribution by an EM algorithm, the robot can discriminate body from non-body by judging which distribution likely causes the variance of sensory data in the current observing posture. Experiments with real robots show the validity of the proposed method.