Multi-Objective POMDPs for Robust AutonomyDownload PDF

Published: 09 Jul 2020, Last Modified: 05 May 2023RSS 2020 Robust Autonomy WorkshopReaders: Everyone
Abstract: Decision-making in real-world robots requires a robustness to uncertainty in dynamic environments with a balancing across multiple objectives. This paper proposes a general model for robust multi-objective reasoning called a topological partially observable Markov decision process (TPOMDP) and its fully observable subclass (TMDP). TPOMDPs and TMDPs allow for additional objective measures, such as maximizing safety, smoothness, and/or other human preferences, to be incorporated into a typical POMDP or MDP objective, such as minimizing time or distance traveled. To enable use on a real robot, we also present a scalable solver for TPOMDPs. The model is discussed through comparisons of POMDP policies on a fully operational autonomous vehicle prototype acting in the real world.
TL;DR: Multi-objective POMDPs can be used to model various robustness objectives. The model admits an topological ordering over these objectives, with slack, allowing for a rich landscape of decision-making concerns to be modeled.
Keywords: MDP, POMDP, Multi-Objective, Autonomy, Robots
1 Reply

Loading