Supervised Learning-Based Hierarchical Local Trajectory Planning Framework for Quadrotors

In the field of quadrotor trajectory planning research, mainstream methods are mainly based on optimization ideas. However, due to their reliance on manually set local optimization iterative rules, these methods may lead to poor overall planning results in complex environments. On the other hand, end-to-end learning-based planning methods face challenges in ensuring flight safety during drone trajectory planning. To address generalization challenges in varying environments, this paper proposes a robust and efficient local planning algorithm framework for drones. A lightweight neural network inspired by environmental perception information is used to predict the future trajectory of the drone, serving as an initial guess to inspire the back-end planner to perform rule-based local obstacle avoidance to ensure the safety of the trajectory. Experiments in a variety of complex simulation environments have verified the superiority of the proposed method in terms of planning efficiency and effectiveness.