Emma Brunskill

Title: Tractable, Approximate POMDP Planning for Robotics

Abstract:
Explicitly representing the partial observability resulting from imperfect
sensors has been one of the key themes in the robotics research
community over the last 20 years.  Many robotic domains,
including navigation, monitoring, and grasping, can be
described as a POMDP. However, these domains often involve
continuous states or actions, and may require long horizon planning
to achieve good performance: in this sense, they embody the
frequently described curses of dimensionality and history. To tackle
robot problems, some researchers have leveraged or assumed
structure in the domain, such as factored state spaces or
Gaussian belief states. Certain other (sometimes overlapping) approaches
have gained tractability by restricting the policy space considered
through the use of macro-actions. I will discuss some encouraging examples
of using such approximate POMDP planners to solve robotics tasks,
including a surveillance monitoring problem using an autonomous
helicopter, and a robot grasping application.