Adaptive sampling with a robotic sensor network

Robotic sensor networks provide new tools for in-situ sensing in challenging settings such as environmental monitoring. Motivated by applications in marine biology we study the field reconstruction problem using a robotic sensor network. We focus on the adaptive sampling problem. The network makes sequential control decisions about where to sample the environment to optimally reconstruct the field being probed. Reconstruction errors are thus reduced by adjusting the sample distribution. In a robotic sensor network, the static network nodes provide long term continuous sensor data (samples) at fixed locations. On the other hand, the mobile nodes (robots) are able to adjust the distribution of the samples but the number of samples they can take is limited. We present algorithms that exploit the advantages of both static and mobile nodes. Samples from static nodes are used to bootstrap a coarse estimate of the field, and the robots take additional samples to successively refine the estimate. Three cases are studied in this dissertation: a static network augmented with (1) a single robot, (2) a large number of robots, and (3) a small number of robots. For the first case, approximation algorithms developed for the orienteering problem are used for generating near-optimal data-gathering tours once the gain of each location is computed by using optimal experimental design for non-parametric regression. For the second case, an auction-based algorithm is used for coordination between mobile nodes. For the final case, spatial and task partitioning methods are used to coordinate robots with each partition being treated as in the single robot case. The methods developed in this dissertation have been extensively tested in simulations. Experimental field trials (several km in two lakes and a harbor) on a physical network of two robotic boats and ten static buoys validate the algorithms.