Imitation is a powerful and versatile mechanism for acquiring new behaviors. Studies involving human infants have revealed that imitation-based learning begins very early in infancy and progresses through a series of stages, culminating in sophisticated forms of imitation such as imitation based on inferring intentions and goals. In this talk, I will describe a model of imitation and sensorimotor learning that relies on Bayesian inference in probabilistic graphical models. The graphical models are used to encode "forward models" of the environment and capture the probabilistic consequences of motor actions. These forward models are used in conjunction with sensory likelihoods and prior distributions over actions and goal states to obtain posterior distributions for action selection. I will discuss the application of such a model to understanding imitation in infants and implementing imitative behaviors in a humanoid robot.
Copyright. All Rights Reserved.