Dynamic Environment Mapping for Autonomous Thermal Soaring

Abstract

This paper describes a map-based, high-level control algorithm for autonomous thermal soaring. The algorithm combines ideas from occupancy grid maps and value iteration algorithms used in the robotics and machine learning communities. Estimates of the specic energy rate throughout the ight domain are built from three sub-function components. The rst component is used to drive exploration when no favorable energy regions are known. A short-term memory component synthesizes recent sensor measurements to map the energy available in the neighborhood of the UAV and enables the detection of and centering in moving thermals. The nal component uses a history of previous atmospheric energy measurements to identify patterns in the ight domain and allow the vehicle to return to locations that consistently form thermals. Seven variations on the high-level control algorithm are tested using a six degree of freedom simulation and compared to bounding cases where the controller either has perfect knowledge of the thermal eld or has no thermal sensing ability and ies a large diameter circle. The objective of interest is to maximize the aircraft’s average energy state over a one hour simulation. Due to the stochastic nature of the thermal model, 100 trials are run for each variation of control algorithm. Simulations indicate that on average the best performing autonomous thermal soaring controller achieves 62% of the possible improvement in the objective. For each algorithm there are large variations in the performance achieved due to the stochastic nature of the thermal elds.