Infant Robot Interaction

Socially Assistive Robotics: increasing frequency of infant leg movements

 

Exploring Bayesian Surprise as a Predictor of Infant Visual Attention

The above video depicts a user study as part of a collaborative effort by the Interaction Lab and Infant Neuromotor Control Lab to encourage more frequent leg movement in infants. In order to keep an infant engaged during the experiment, the robot must be able to reliably capture the infant’s attention. Understanding which factors can predict infant attention may help us to design more salient robot behaviors. In this project, I evaluated whether a Bayesian model of low-level surprise, developed by Laurent Itti and Pierre Baldi, could predict infant attention, finding that infants in the study looked more often at areas of their environment that had a higher level of Bayesian surprise. This work was published in the 28th IEEE International Conference on Robot & Human Interactive Communication (pdf).

Video footage of the experiment from the infant’s point of view. The target represents the gaze location of the infant. The head-mounted camera and gaze location tracker are products of Positive Science.

Video footage of the experiment from the infant’s point of view. The target represents the gaze location of the infant. The head-mounted camera and gaze location tracker are products of Positive Science.

Bayesian Surprise values at 16x16 pixel patches of the video from the head-mounted camera. Light values represent high levels of surprise, while dark values represent low levels of surprise.

Bayesian Surprise values at 16x16 pixel patches of the video from the head-mounted camera. Light values represent high levels of surprise, while dark values represent low levels of surprise.