부산대학교 도서관

Real-time path planning typically aims to obtain a collision-free and shorter path with lower computational complexity for UAVs in unknown environment. Apart from the above basic objective, kinematic constraints and the smoothness of path should be further considered especially for fixed-wing UAVs restricted by their maneuverability. In this paper, we propose an adaptive real-time path planning method based on Deep Reinforcement Learning. Taking the sensor data of obstacles nearby and the target’s position relative to the UAV as the decision information, and designing the action satisfying kinematic constraints of fixed-wing UAV, the proposed method can plan a feasible path for fixed-wing UAV in real-time. Experimental results show that the adaptive action devised combining with greedy reward, granularity reward and smoothness reward can accelerate the convergence speed of the algorithm and enhance the smoothness of the planned path.