Machine Learning and Friends Lunch
Reasoning about sensor uncertainty in probabilistic planning
Brendan Burns
UMass
Abstract
Sampling-based motion planning algorithms have significantly
improved
the state of the art in robotic motion planning. However these
improvements
have been based upon the assumption that knowledge of the
robot's workspace is perfect. When sensors are used to model
the
environment they introduce noise and error. Motion planners
must be
aware of this error to compute plans that minimize the
probability of
failure. I'll describe recent work developing efforts to
incorporate
reasoning about uncertainty and the source of perceptual error
into planning
algorithms.