loader graphic

Loading content ...

Data-driven Learning and Modeling of AUV Operational Characteristics for Optimal Path Planning

Edwards, J., J. Smith, A. Girard, D. Wickman, D. N. Subramani, C. S. Kulkarni, P.J. Haley, Jr., C. Mirabito, S. Jana, P. F. J. Lermusiaux, 2017. Data-driven Learning and Modeling of AUV Operational Characteristics for Optimal Path Planning. In: Oceans '17 MTS/IEEE Aberdeen, 19-22 June 2017, In press.

The utilization of Autonomous Underwater Vehicles (AUVs) such as propelled vehicles, gliders, and floats is rapidly growing for a wide range of missions and ocean regions. For optimized utilization, the operational characteristics of the AUVs need to be modeled as accurately as needed by the optimization and specific needs of the ocean missions considered. The advent of machine learning and data sciences provides an opportunity to augment the classic engineering modeling and laboratory analyses by learning the AUV operational characteristics in situ, during and after each sea operations. Such data-driven learning is critical because, from mission to mission, the AUV usage frequently differs, the dynamic ocean environment changes, and the configuration of the AUV itself changes. For the latter, considering propelled vehicles, it is for example very common for fins and buoyancy to be modified, for payloads to be changed, and for the internal content and overall body of the AUVs to be altered. We illustrated the use of in-situ-data-driven learning and modeling of operational characteristics of AUVs for path planning. The operations and learning experiments were conducted in the Buzzards Bay, Vineyard Sound, and Martha Vineyard’s region for several AUV configurations, missions, and ocean conditions. Specifically, we identified and applied simple methods to estimate the relationships between thruster RPM with forward vehicle speed and to confirm that the specific fin configuration affects the net forward speed of the REMUS 600. Such data-based learning should be completed in real-time so as to ensure accurate F(t) models and thus time-optimal performances. These results can be employed for other types of optimal path planning and AUV missions, including energy, sensing, and surveillance optimality.