Path Planning for Fixed-Wing Unmanned Aerial Vehicles
Schneider, Daniel · Repository for Publications and Research Data (ETH Zurich) · 2016
This thesis presents a first implementation of a real-time capable onboard path planning framework computing shortest paths for fixed-wing aerial vehicles for many start-goal configurations.The framework provides kinodynamic planning based on the Dubins airplane motion model.The framework is implemented as a ROS node what makes it executable on common companion-computers installed on fixed-wing aerial vehicles.We evaluated 8 different sampling-based motion planning algorithms in theoretical and real experimental setups.Two new ideas to reduce the time for finding shortest paths for the Dubins airplane are explained and compared against existing planning methods.On the one hand, the informed subset of the Dubins airplane for minimum path length is approximated as a subset of the informed subset for a system without differential constraints.On the other hand, in order to find initial paths as fast as possible, the initial path is planned using straight-line connections between samples instead of dynamically feasible paths.A modification to the optimal Fast Marching Tree algorithm (FMT*) is presented.The modification allows to predict the amount of samples workable in the time available for planning and thereby makes FMT* usable for real-time and onboard path planning.Experiments have shown that planning initial paths with straight lines and approximating the informed subset for the Dubins airplane for minimum path length allows finding initial paths in less than 1 second and expedites the convergence to the optimal path such that after 2 seconds reasonably short paths are found.Furthermore, planners iteratively drawing one sample at a time, e.g.optimal Rapidly-exploring Random Tree (RRT*), perform better than planners drawing batches of samples, e.g.Batch Informed Tree (BIT*), if planning in wide open spaces.For cluttered, narrow and twisted maps, it is the other way round.Additional experiments have shown that computing non-optimal paths for start-goal configurations occurring rarely in real application scenarios, speeds up the computation of the paths up to 40 times.Further experiments have shown that using a light collision checking algorithm reduces the time spent for collision checking more than 10 times.iv