Planning for perception and perceiving for decision: POMDP-like online optimization in large complex robotics missions

Abstract

This ongoing phD work aims at proposing a unified framework to optimize both perception and task planning using extended Partially Observable Markov Decision Processes (POMDPs). Targeted applications are large complex aerial robotics missions where the problem is too large to be solved off-line, and acquiring information about the environment is as important as achieving some symbolic goals. Challenges of this work include: (1) optimizing a dual objective in a single decision-theoretic framework, i.e. environment perception and goal achievement ; (2) properly dealing with action preconditions on belief states in order to guarantee safety constraints or physical limitations, what is crucial in aerial robotics ; (3) modeling the symbolic output of image processing algorithms as input of the POMDP's observation function ; (4) parallel optimization and execution of POMDP policies in constrained time. A global view of each of these topics are presented, as well as some ongoing experimental results.