Target Tracking for Quadrotors based on Deep Reinforcement Learning
Yan Gao, Feiqiang Lin, Changyun Wei, Raphael Grech, Ze Ji · 2024
In this paper, we propose a deep reinforcement learning-based method for quadrotors to learn depth-based tracking policies autonomously. To this end, we present a novel reward function that guides the quadrotor to follow the target, avoid collisions, and keep the target close to the centre of the onboard camera’s field of view without occlusions. In addition, to improve learning efficiency, we suggest using a teacher-student learning strategy. Specifically, we first train a state-based teacher policy encoding low-dimensional obstacle information, which then guides the vision-based student policy during training. Moreover, we introduce a variant of the Proximal Policy Optimisation algorithm based on the importance sampling algorithm. It facilitates the teacher-student learning process and enables the vision-based agent to escape local minima. The experimental results have demonstrated the satisfactory performance of our proposed method.