Bootstrapping Imitation via Vision Based Shared Attention

by
Matt Hoffman

Shared attention refers to the simultaneous perceptual focus of two or more agents on a single object in their shared environment. Endowing robots with the capacity for shared attention can lead to systems capable of complex, natural forms of learning, such as learning by imitation. We discuss various methods of building shared attention on a robotic platform. Our methods are based on Meltzoff and Moore's AIM model for imitation in infants. Bayesian algorithms implement the core of a shared attention framework that follows the gaze and finger-pointing of an instructor and learns task-specific saliency models from a cluttered scene. Our results demonstrate the value of our system for promoting interaction between humans and robots.

Advised by Raj Rao

CSE 403
Wednesday
May 11, 2005
3:30 - 4:20 pm