Toward Model Free Atmospheric Sensing by Aerial Robot Networks in Strong Wind Fields Jack Elston * , Maciej Stachura * , Eric W Frew * and Ute C. Herzfeld † * Aerospace Engineering Sciences University of Colorado, Boulder, CO 80309 † Cooperative Institute for Research in Environmental Sciences University of Colorado, Boulder, CO 80309-0449 Abstract—This paper presents a system for in situ atmo- spheric sensing using an aerial robot system in the presence of a strong wind field. The geostatistical concept of the variogram is used to characterize regions of the environment with high variability, which is assumed to correlate with scientific interest. After regions of interest are identified, ordered upwind methods are used to generate feasible trajectories in the face of strong background wind. The feasible trajectories are combined with the variogram characterization to select feasible paths that travel through the regions of highest interest. The system is tested in simulation using data from a simulated severe storm. I. I NTRODUCTION Atmospheric sensing by aerial robot networks has the potential to yield unprecedented types and amounts of data for a large range of scientific applications. Unmanned aircraft systems (UAS) have already been fielded for missions such as pollutant studies [1], polar weather monitoring [2], and hurricane observation [3]. Proposed UAS span numerous future applications that include severe storm penetration [4] and persistent global climate observation [5]. In most of these applications, the nature of the problem leads to several constraints on the planning and control ar- chitecture that can be applied. First, path planning algorithms cannot rely on high-fidelity models of the phenomena of interest. These types of models can be too computationally intensive to include in a real-time planning loop and they often have poor accuracies since the primary reason for studying the phenomena is to characterize them at certain locations and resolutions. Second, the phenomena of interest have to develop significantly before they can be identified. This means that deployment of the UAS will have to be made with limited time for ingress and data collection. Third, atmospheric phenomena can be highly dynamic so a large amount of data must be collected from a fairly sizeable area in a short amount of time in order to inform the modeling process. Finally, nonlinear aircraft dynamics and the presence of strong wind fields (fields that contain wind speeds larger than the speed of the aircraft) complicate path planning for the UAS. Planning and control to optimize sensing has been applied to a large number of problems. This work can be broadly placed into four categories. The first is driven by model based estimation and relies on the calculation of information measures that are functions of input states [6], [7]. These methods can be applied to dynamical systems described by partial different equations (e.g., most atmospheric phenom- ena), although the number of states involved in the simulation of complex phenomena make it difficult to generate inputs in real time. The second category, often referred to as adaptive sampling, is driven by the directed sampling of a scalar field in order to obtain statistical properties or to locate sources, contours, and boundaries [8], [9]. The final two categories are both referred to as coverage control. The first approach deploys vehicles through nearest neighbor laws to achieve an optimal deployment of a sensor team based on an underlying set of basis functions [10], [11]. Here, there is an assumption that robots can communicate in order to calculate relative positions and achieve consensus on the underlying functions. Alternatively, coverage control refers to distributed search or patrol of an area such that every point in the environment is “covered” by a sensor to a given level of quality [12]. The architecture presented here will employ a new form of adaptive sampling based on spatio-temporal statistics of the sampled atmospheric phenomenon. For the atmospheric science applications described here the end user scientists need in situ data in order to develop models. Hence, the model based approaches do not apply. Likewise, the notion of a simple contour or gradient to follow may not exist. Further, sensing can only occur locally so optimal deployment algo- rithms ignore large portions of the environment. Likewise, the phenomena of interest cannot be described by a simple set of basis functions that can be estimated so the first form of coverage control does not apply. Finally, time constraints prohibit detailed sensing of the entire environment, so the second coverage control approach is also inappropriate. Two challenges for the efficient sampling of atmospheric phenomena are addressed by the architecture described here. The first is identification of regions of interest within the phenomenon where resources should be directed. The geo- statistical concept of the variogram [13] is used to derive a feature vector that characterizes the expected value of more sampling in a region. The second is path planning in the presence of strong background winds. The planning problems considered here are examples of control affine systems with drift. In general, techniques such as Lie algebra can be extended to determine controllability of systems with drift though optimal motion planning for them can be extremely