基于深度强化学习的无人机自主导航应用研究

In order to solve the problems of traditional UAV autonomous navigation algorithms, such as low accuracy of the implementation and the consumption of extensive computational resources for constructing high-precision maps, a new navigation algorithm based on deep learning and reinforcement learning methods was proposed, with end-to-end implementation and no need to construct map a priori. Firstly, a high-fidelity and high-performance simulation environment is designed and built based on Airsim and UE4 platforms to train UAV\'s autonomous navigation tasks. It supports domain randomization and gym interface to achieve a photo-realistic level of information quality. Secondly, using reinforcement learning ideas, a network model for autonomous UAV navigation tasks is designed and trained by designing the state and action space and formulating the reward function. A multi-actor-single-critic PPO algorithm network structure incorporating an attention mechanism was designed by optimizing the problem of over-simplified scenarios for single-actor in an actor-critic framework-based PPO algorithm. Finally, simulation experiments verify the method\'s feasibility and effectiveness.