An MDP-based approximation method for goal constrained multi-mav planning under action uncertainty

Abstract

This paper presents a fast approximate multi-agent decision theoretic planning method extended from the well-known Markov Decision Process (MDP). Our objective is to plan motions for a team of homogeneous micro air vehicles (MAVs) toward a set of goals, such that each MAV at any state at any moment follows an action policy toward a unique goal, while considering action uncertainty. We pursue an efficient formulation by first considering a deterministic abstraction of the stochastic system based on approximate initial paths. These deterministic and decoupled sub-problems are converted to the stochastic domain and improved by individual agents or a subset of agents. The resulting decoupled formulation requires processing of a partial state space and enables online operation given applications with emerging tasks.