HIGH LEVEL ROBOT MOTION PLANNING USING POMDP

Abstract

Motion planning in uncertain environments is an essential feature of autonomous robots. Partially observable Markov decision processes (POMDP) provides a principled approach for such planning tasks and have been successfully employed for various robotic applications. Offline planning algorithms for POMDPs have proved to achieve optimal policies. However, these algorithms are computationally very expensive and are not often applicable to accomplish realistic robot motion planning scenarios. As an alternative to offline planning Monte Carlo algorithms for online planning were developed, e.g. DESPOT and POMCP are implemented for public use within a set of open-source POMDP libraries. The aim of this master thesis is to explore applicability of available open-source POMDP libraries to real-time robot motion planning tasks. We adopt these libraries for the POMDP models proposed in the literature and replicate a number of realistic benchmark experiments.