Gaussian belief space planning using probabilistic graphical models

Co-supervisor: Corné van Daalen

To successfully perform a user-specified task, an autonomous robot must plan its future actions using predictions of the (continuous) robot and environment states. When the robot cannot observe the states directly, it represents its knowledge of the states as a belief distribution. The goal of belief space planning is then to find a policy (or mapping from belief to actions) that minimises a cost function corresponding to the specified task. Previous research efforts towards planning (e.g., motion planning and Markov decision processes) were successful in addressing either uncertainty or continuous state spaces, but not both.

A promising way to model the problem of decision-making under uncertainty is using influence diagrams – a specific type of probabilistic graphical model (PGM). PGMs are a family of statistical techniques that represent the local structure of an inference problem using directed or undirected graphs. In addition, PGMs make it easy to extend a model or make approximations. The aim of this project is to investigate the application of PGMs to the belief space planning problem. The use of Gaussian distributions would further allow planning for systems with continuous dynamics, without the traditional use of discretisation. This will aid planning for tasks like mapping an environment or avoiding collisions with static or dynamic obstacles. The experiments for this project will be performed in simulation only.