Performance evaluation of decision making under uncertainty for low power heterogeneous platforms

Abstract

Value Iteration (VI) is a core method to find optimal policies, allowing a robot to act autonomously in environments where the effects of its actions are not deterministic. Although there are extensive studies on VI’s theoretical properties and computational cost, its energy performance — an essential indicator for its use in the physical world — has not been evaluated. In this paper, we explore both the energy and runtime performance of five parallel implementation strategies of VI on representative low-power heterogeneous computing platforms that integrate CPUs and GPUs, for the use-case scenario of indoor autonomous robot navigation. We provide a statistical analysis of their performance depending on the problem size, parallel implementation and computing platform. Our study shows that CPU–GPU heterogeneous strategies reduce computation time and energy considerably, given large enough problem sizes, regardless of the computation platform. This work also provides practical guidelines to assist in the application of the most efficient implementation, either in terms of energy consumption or time, to a low-power heterogeneous platform.