Monday, February 22, 1999
3:00 PM (refreshments 2:45)
Room NE43-518
EECS Special Seminar
Abstract
In robotics, it is traditional to build metric, Cartesian maps of the environment. Such maps hard to build and, I'll argue, are often unnecessarily precise. In this talk, I'll describe another approach in which we represent maps as partially observable Markov decision processes (POMDPs), consisting of a discrete set of states and stochastic transition and observation models. Such maps are easier to learn and support POMDP planning techniques, yielding robust behavior in highly uncertain and ambiguous environments. It turns out, however, that a certain amount of metric information is crucial for learning. I will conclude by discussing a new project that will explore vision-based map-learning in humans and robots.
|
Modified: Feb 18, 1999|
Current events|
Your comments
and inquiries are welcome.