In context
As low-altitude airspace opens up and the low-altitude economy expands, unmanned aerial vehicles and other low-altitude aircraft are increasingly used for logistics, monitoring, and emergency response. However, their complex operating environments pose safety challenges, especially with unknown threats. Existing path planning methods—classical algorithms like A* and RRT, and swarm intelligence methods—struggle with dynamic unknown threats or can fall into local optima. Deep reinforcement learning offers adaptability but suffers from low training efficiency and slow convergence.
What was reported
Researchers from Nanjing University of Aeronautics and Astronautics proposed a TCP-DQN (target-guided curriculum learning and prioritized replay deep Q-network) algorithm for dynamic path planning of low-altitude aircraft in complex 3D environments. The method integrates curriculum learning into the reinforcement learning framework, setting up environments of increasing difficulty (low, medium, high) to accelerate training. A target-guided maneuver strategy is used to improve path feasibility, while a combined reward function addresses sparse reward issues. Prioritized replay of obstacle avoidance experiences enhances learning performance.
The action space is simplified to nine discrete maneuvers (e.g., fly forward, up-left, down-right) to balance precision and training efficiency. The cost function includes length, height, collision, and threat costs, with constraints on yaw angle, pitch angle, and flight range. Simulations in dynamic unknown threat environments demonstrated that TCP-DQN can quickly plan safe and efficient paths, improving training speed and convergence compared to standard DQN.
Why it mattered
This work advances deep reinforcement learning for autonomous navigation in low-altitude aircraft, addressing key limitations of training efficiency and path feasibility. The curriculum learning and prioritized replay techniques could be applied to other robotic path planning tasks, enhancing the deployment of autonomous systems in dynamic, unknown industrial environments.
The simulation results demonstrate that the algorithm can quickly plan the safe and efficient paths for low-altitude aircraft in dynamic and unknown threat environments.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2025-06-09 · “复杂环境下基于 TCP-DQN 算法的低空飞行器动态航路规划”