The general notion of a Perceptive Agent is quite broad, and there are several challenges that it raises. In this talk I will focus in particular on one key problem, how to construct an attentive interface that overcomes the inherent limits of fixed sensors. I will present a model of attention which learns a policy for perceptual actions from experience, via a gesture recognition task. My model uses a hidden-state learning paradigm, based on Q-learning and the Partially Observable Markov Decision Process. To solve a multiple target recognition problem, I use a novel extension of the Q-learning framework to handle vector-valued rewards and utility values.
Finally, I will show two applications of Perceptive Agents, one for interacting with computer-generated graphical worlds (our ALIVE system of perceptually-situated virtual characters), and another for telepresence, animating realistic Avatars which match the body pose and facial expression of a user in real-time.
HOST: Prof. G. Sussman
|
Modified: Jun 25, 1997|
Current events|
Your comments
and inquiries are welcome.