Multi-timescale Sensor Fusion and Control

Networked autonomous systems are a rapidly expanding area of research and development across academic, commercial, and military endeavors. Significant challenges exist in extending traditional detection and estimation methods to such distributed systems of sensors when we relax assumptions on full communications connectivity and global observability of the network. Global observability can be interpreted as a persistent coverage of all degrees of freedom associated with a object's feature vector - this can be satisfied by a combination of physical diversity of homogeneous sensors and/or diversity across sensing domains for heterogeneous sensors, and the role of resource allocation across the network is to determine configurations, and reconfigurations, of platforms that achieve said diversity. In a general heterogeneous sensor network, persistent global observability across the entire area of operations requires control decisions at a much longer timescale than the feature estimate updates that provide locally full rank observability. In this paper, we temporally separate the long-timescale resource allocation control process from the parameter estimation through the use of a decentralized Partially Observable Markov Decision Process (POMDP) control model that employs consensus estimates on object features as observations and benchmark this multi-timescale approach against centralized Linear Quadratic Gaussian (LQG) control for a fully connected network with simultaneous estimation and control updates.