Gaussian Processes for Plume Distribution Estimation with UAVs

Recent scientific work explored the possibility to use mobile robots for environmental monitoring. This includes for example the estimation of ozone concentrations or locating the source of a pollutant plume. So far the modeling of the complete spatial distribution of a plume (which has different spatial characteristics compared to the ozone concentrations) has not been done. In this work existing methods of Bayesian optimization, namely global optimization (GO) and the distance based upper confidence bound (DUCB), were evaluated on this task. Also, a new method – plume distance based upper confidence bound (PDUCB) – and an extension to multiple robots is proposed. All methods were tested in simulations using the QRSim quadrotor simulator. The existing methods were not able to solve the task satisfyingly, whereas the PDUCB method was able to approximate plume distributions with noisy measurements reasonably well.