Path Planning and Flight Control of Drones for Autonomous Pollination

The decline of natural pollinators necessitates the development of novel pollination technologies. In this thesis, a drone-enabled autonomous pollination system (APS) that consists of five primary modules: environment sensing, flower perception, path planning, flight control, and pollination mechanisms is proposed. These modules are highly dependent upon each other, with each module relying on inputs from the other modules. This thesis focuses on approaches to the path planning and flight control modules. Flower perception is briefly demonstrated developing a map of flowers using results from previous work. With that map of flowers, APS path planning is defined as a variant of the Travelling Salesman Problem (TSP). Two path planning approaches are compared based on mixed-integer programming (MIP) and genetic algorithms (GA), respectively. The GA approach is chosen as the superior approach due to the vast computational savings with negligible loss of optimality. This path planning approach is applied to 2D and 3D APS missions, pollinating both row crops and crops grown on trees, such as strawberries and almonds respectively. To accurately follow the generated path for pollination, a convex optimization approach is developed to solve the quadrotor flight control problem (QFCP). This approach solves two convex problems. The first problem is a convexified three degree-of-freedom QFCP. The solution to this problem is used as an initial guess to the second convex problem, which is a linearized six degree-of-freedom QFCP. It is found that changing the objective of the second convex problem to minimize the deviation from the initial guess provides improved physical feasibility and solutions similar to a general-purpose optimizer. A method of sparse environment collision avoidance using the convex approach and the stitching together of multiple control sequences is introduced. The path planning and flight control approaches are then tested within a model predictive control (MPC) framework where significant computational savings and embedded adjustments to uncertainty are observed. Coupling the two modules together provides a simple demonstration of how the entire APS will operate in practice.

This is the third preliminary submission, with corrections made to formatting.