Design and simulation of a machine learning-based approach for an autonomous UAV for use in Agriculture 4.0 applications

The aim of this thesis is to design, build and simulate an algorithm which provides guidance and control for autonomous UAVs in Agriculture 4.0 within the PRIN project “New technical and operative solutions for the use of drones in Agriculture 4.0”. The simulation scenario includes several targets which can represent plants, vegetables, or other vegetation where treatments deployed by the UAV is needed. The main objective of the developed algorithm is to guide the UAV, identifying and reaching all possible targets autonomously using information derived by depth cameras and other sensors. Another objective of the project is the application of the final algorithm with limited GPS or noisy signal conditions. This, in fact, is one of the most critical working scenarios and traditional guidance based on mission definition through waypoints could fail to achieve an acceptable level of positioning precision depending on the UAV’s mission profile. The simulation environment is developed using ROS (Robot Operating System) framework and the Gazebo simulator. Simulation includes the UAV and a physical environment which should represent a working scenario as close as possible to the real one. During the project will be developed an allocation algorithm, which objective is to map target’s positions in the local frame of reference, and a control method for three axes. Also, two Machine Learning (ML) applications are part of main challenge of the project. A U-Net, deep neural network, for target recognition will be implemented and a Reinforcement Learning (RL) algorithm, that describes the policy with which targets to reach sequentially are chosen, will be tested. Final results show the algorithm works in the simulated scenario, with the UAV being able to allocate each target in the correct position in local frame and subsequently reach them one after another without passing over the ones already reached. This is the case if a policy of following always the nearest target is passed to the UAV.?? The Reinforcement Learning method developed presents good but not optimal results: the UAV is able to complete the mission without hovering over the same target twice, but the algorithm will not learn to choose the best path, i.e. the one which requires less time. Some further testing in this case are required or heavy modification of the algorithm are needed.