Mobile robots as remote sensor platforms for characterizing the distribution of vegetation in arid environments

Paul Reverdy 1, Scott Van Pelt2, Ted Zobeck3, Daniel Koditschek1
1University of Pennsylvania, Philadelphia, PA, USA, 2USDA, Agricultural Research Service, Wind Erosion and Water Conservation Research Unit, Big Spring, TX, USA, 3USDA, Agricultural Research Service, Wind Erosion and Water Conservation Research Unit, Lubbock, TX, USA

In arid environments, vegetation and aeolian processes interact in complex ways. To understand either requires integrated measurement of both. In the literature, the distribution of vegetation, e.g., coppice dunes, has historically been characterized at large length scales by remote imaging studies, which can be heavily automated, and at small length scales by observational studies, which reveal more small-scale features but are labor intensive. Aeolian processes have been characterized by a variety of methods, including fixed stations equipped with various sensors.

Our work aims to develop mobile robots that can be used to measure vegetation and aeolian processes simultaneously. In contrast to current vegetation measurement practices, mobile robots equipped with appropriate sensors could provide small-scale measurements with relatively little labor. Furthermore, by carrying relevant sensor payloads, robots can characterize aeolian processes occurring near vegetation, thereby producing novel data sets relevant to better characterizing their interactions. The robots' mobility would also allow them to be dynamically repositioned to maximize the scientific value of the data being collected.

Existing mobile robots are beginning to have the capability to locomote in complex environments such as dune fields. Therefore, our work focuses on developing algorithms to allow such robots to take the relevant measurements in a partially automated manner to complement human labor. As a first step in the algorithm development process, we focus on characterizing the distribution of vegetation in the neighborhood of a given location. We assume the existence of sensors that can detect vegetation within their range; such sensors can be developed by applying appropriate signal processing techniques to existing devices such as lidar scanners. We adopt the following method: have the robot cover a sufficiently large section of terrain in the neighborhood of the starting location, detecting and registering the locations of vegetation as it travels. Three central problems must now be addressed: 1) efficiently covering the terrain, 2) deciding what is a sufficiently large section, and 3) avoiding driving through the individual plants, which constitute obstacles to the robot's motion.

We develop our method by modeling the locations of individual plants as being drawn from a Poisson point process with fixed intensity parameter. Under this model, we show that driving the robot in a spiral pattern leads to efficient sampling and develop a stopping criterion that allows the robot to decide when it has covered a sufficiently large section of terrain. Finally, we develop an algorithm to implement the spiral coverage pattern while avoiding obstacles such as plants. We demonstrate the algorithm in simulation and describe the steps required to implement it on existing mobile robots.