Real-time path planning of controllable UAV by subgoals using goal-conditioned reinforcement learning

The conventional path planning problem for an unmanned aerial vehicle (UAV) typically involves a pre-defined environment and mission, with the objective of reaching a single target point. However, in order to perform different missions, the agent must be trained from scratch. In this paper, we propose a new path planning algorithm for UAVs by training them to be controlled by subgoals, which enhances their degree of freedom to perform various maneuvers. The subgoals can be defined by the user and given to the agent in real-time, allowing the UAV to perform diverse flight missions without prior knowledge of the environment. To achieve this, we utilize goal-conditioned reinforcement learning to train the UAV agent to reach various goals by learning different flight maneuvers. In experiments, we designed specific scenarios to test the UAV agent's ability to perform concrete missions, such as high-flying, low-flying, penetrating, and bypassing. The experimental results show that the same UAV agent trained in a simple environment can accomplish difficult missions in various scenarios. The pre-trained UAV agent can be utilized in other environments as it can be controlled by the subgoals.