Efficient sensing using mobile robots --- From theoretical foundations to in-field validations

Many environment monitoring applications, such as water quality monitoring in rivers and lakes and search and rescue in a disaster struck environment, involve observing uncertain environments with complex spatial and temporal dynamics. We developed an iterative experimental design methodology to understand the spatiotemporal dynamics in such uncertain environments with high resolution. This iterative methodology involves adapting the experimental design in-field, based on expert opinion, prior understanding, and recent observations made during previous experiments. However, efficient monitoring of complex environments require models known to most accurately represent the uncertain dynamics, together with the efficient experimental design methodology. We extend the detailed understanding from iterative, in-field characterization to provide a generic approach for creating a class of stochastic space time models.;To overcome the prohibitive complexity of efficient monitoring of these complex environments using only a network of static sensors, we use actuated sensors---mobile robots with attached sensors that can be moved across the spatial expanse of the observed environment. Typically, these robots have bounded resources, such as limited battery or limited time to obtain measurements. Thus, careful coordination of their paths is required in order to maximize the amount of information collected by them, while respecting the resource constraints. We provide multiple efficient algorithms, with provably strong approximation guarantees for near-optimally solving the NP-hard optimization problem of planning such informative paths. The algorithmic guarantees hold true in the non-adaptive setting, where the models that represent the observed environments are assumed to be known accurately a priori.;Finally, while performing path planning with constrained resource, the most complex problem is to trade off exploration (gathering information about the environment and refining the model used to represent the complex environment) and exploitation (using the most recent model of the observed environment most effectively) to maximize the collected information. We extend our proposed algorithms for the non-adaptive setting to perform adaptive informative path planning that addresses the exploration-exploitation tradeoff, updating the model every time a new observation is made. We further extend our single robot approximation algorithms for both the adaptive and non-adaptive setting to perform path planning for multiple robots while maintaining the strong approximation guarantee provided by the single robot approximation algorithms.;We provided extensive empirical evaluation for our proposed algorithms using the real world sensing datasets including the data collected for scientific studies at lake and river environments. We also validated the effectiveness of our proposed algorithms and their practical applicability through several experiments performed in-field. Data collected during these in-field deployments had been used for multiple scientific studies to further advance the understanding of these complex environments.