ERI: AQUABOT: Advancing Water Quality Monitoring with an Intelligent Bot via Integrated Learning and Decision-Making

Despite substantial technological progress in recent years, autonomous surface vehicles (ASVs) for water quality mapping, such as measuring contaminants, salinity, pH, dissolved oxygen, and temperature, continue to face persistent challenges. Existing ASV planners neither provide calibrated uncertainty estimates that separate sensor noise from model uncertainty nor incorporate operational realities such as collision risks, energy limitations, and mission time constraints. This Engineering Research Initiation (ERI) award supports fundamental research that holistically tackles these interconnected challenges through an iterative process of sensing, modeling, and learning, enabling progressively better uncertainty quantification and more informed decision-making in large-scale aquatic environments. The proposed framework is designed for broad applicability across aquatic environments, from freshwater quality assessment to marine ecosystem monitoring, ensuring sustained impact across multiple disciplines. For instance, results from this work are expected to contribute, directly or indirectly, to pollution control and aquatic ecosystem health, since each of these depends on spatially explicit, timely data that intelligent robotic platforms can provide. Additionally, this award will provide opportunities to educate and train graduate and undergraduate students in the cross-disciplinary areas of robotics, AI, geology, and biology, as well as to collaborate with state and federal agencies and robotics communities through open-source dissemination of data and software. Information-theoretic ASV planning frequently accommodates inadequate uncertainty quantification, neglecting crucial operational constraints such as energy limits and computational constraints, resulting in suboptimal or infeasible trajectories. Therefore, intelligent sensing over large-scale aquatic environments often yields misleading water quality maps. Although research consistently shows that uncertainty quantification in probabilistic models is key to the sequential sampling decisions that build an accurate water quality map, how to separate sensor uncertainty from model uncertainty remains unclear. A closed-loop autonomous monitoring system demands an accurate, uncertainty-aware aquatic model to actively sense water quality parameters and select next-best sensing locations. To accurately quantify uncertainties, a Deep Evidential Regression (DER) model will be developed to enable joint modeling of aleatoric (sensor) and epistemic (model) uncertainties. An information-theoretic planner will enable sequential decision-making over water sampling locations while respecting motion and sensing constraints. Field deployment will be performed at New Orleans’s East Landbridge wetland area to measure water quality parameters such as salinity, pH, dissolved oxygen, and temperature. Several water quality parameter maps will be generated with well-calibrated uncertainty quantification, enabling reliable, long-endurance performance in dynamic and uncertain aquatic environments and laying the foundation for scalable autonomous systems capable of adaptive learning, uncertainty-aware reasoning, and minimal human intervention under challenging field conditions. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.