Multi-robot active sensing of non-stationary Gaussian process-based environmental phenomena

Abstract

A key challenge of environmental sensing and monitoring is that of sensing, modeling, and predicting large-scale, spatially correlated environmental phenomena, especially when they are unknown and non-stationary. This paper presents a decentralized multi-robot active sensing (DEC-MAS) algorithm that can efficiently coordinate the exploration of multiple robots to gather the most informative observations for predicting an unknown, non-stationary phenomenon. By modeling the phenomenon using a Dirichlet process mixture of Gaussian processes (DPM-GPs), our work here is novel in demonstrating how DPM-GPs and its structural properties can be exploited to (a) formalize an active sensing criterion that trades off between gathering the most informative observations for estimating the unknown, non-stationary spatial correlation structure vs. that for predicting the phenomenon given the current, imprecise estimate of the correlation structure, and (b) support efficient decentralized coordination. We also provide a theoretical performance guarantee for DEC-MAS and analyze its time complexity. We empirically demonstrate using two real-world datasets that DEC-MAS outperforms state-of-the-art MAS algorithms.