This work presents a conceptual framework, called FloatCast, for the control of a small fleet of buoyancy-controlled ocean profiling floats. The control objective is to maximize sampling coverage in a given region of interest. The framework optimizes park depth and park duration commands for each float in the fleet. FloatCast uses an Echo State Network to make a sea level anomaly forecast, which is converted into a surface flow forecast. This flow forecast informs a Lagrangian particle model of drifting vehicle dynamics. The state-space model of the float dynamics uses candidate sets of commands to predict float trajectories, which are evaluated using a mapping error scoring metric. Stochastic analysis illustrates a risk-reward tradeoff between uncertainty and potential coverage for candidate float commands. This paper introduces each of these components of FloatCast and presents initial simulation results using float data from a deployment in the Gulf Stream from July 2024.

Over the last twenty years, profiling floats have revolutionized ocean observations with globally distributed Lagrangian arrays performing fixed vertical sampling cycles. Here we investigate adaptive sampling with an array of inter-dependent floats guided by a software package called FlowPilot, which uses all available float measurements to select park depths that provide favorable drifts based on sampling goals. Drift predictions are performed with multiple prediction methods, including methods that use float data (drift velocity, geostrophic velocity calculations) or from external sources like numerical ocean forecast models. A skill-based weight is assigned to each method based on how accurately it predicts recent drifts. With this generalized approach to prediction, disparate methods can be combined numerically to permit multi-method optimization. The emergent skill of FlowPilot is tested and quantified by numerical simulations that minimize dispersion by keeping a grid of floats close to the center of the deployment box.

Buoyancy-controlled underwater floats have produced a wealth of in situ observational data from the open ocean. When deployed in large numbers, or "distributed arrays," floats offer a unique capacity to concurrently map 3D fields of critical environmental variables, such as currents, temperatures, and dissolved oxygen. This sensing paradigm is equally relevant in coastal waters, yet it remains underutilized due to economic and technical limitations of existing platforms. To address this gap, we developed an array of 25 μFloats that can actuate vertically in the water column by controlling their buoyancy, but are otherwise Lagrangian. Underwater positioning is achieved by acoustic localization using low-bandwidth communication with GPS-equipped surface buoys. The µFloat features a high-volume buoyancy engine that provides a 9% density change, enabling automatic ballasting and vertical control from fresh to salt water (~3% density change) with reserve capacity for external sensors.

In this paper, we present design specifications and field benchmarks for buoyancy control and acoustic localization accuracy. Results demonstrate depth-holding accuracy within ±0.2 m of target depth in quiescent flow and ±0.5 m in energetic flows. Underwater localization is accurate to within ±5 m during periods with sufficient connectivity, with degradation in performance resulting from adverse sound speed gradients and unfavorable spatial distributions of surface buoys. Support for auxiliary sensors (<10% float volume) without
additional control tuning is also demonstrated. Overall performance is discussed in the context of potential use cases and demonstrated in a first-ever array-based three-dimensional survey of tidal currents.