UAV logistics path planning in complex constrained environments based on the VT-ITD3 algorithm

Abstract With the widespread adoption of unmanned aerial vehicles (UAVs) in logistics applications, efficiently accomplishing logistics tasks in complex constrained environments has become a prominent research focus. Path planning serves as a critical component in this context. However, achieving reliable navigation in complex environments while respecting practical kinematic constraints remains nontrivial for both classical planners and learning based approaches. To address this challenge, this paper proposes VT-ITD3, a path planning method that integrates artificial potential field (APF) guidance with an improved twin delayed deep deterministic policy gradient (TD3) framework. We introduce a dynamic virtual target generation mechanism that provides adaptive intermediate targets when the direct route to the goal is blocked, which alleviates local minima and improves navigation reliability in complex environments. Moreover, kinematic feasibility is explicitly considered through constraints on velocity and heading change. To alleviate reward sparsity, a composite dense reward provides stepwise feedback. Furthermore, an improved prioritized experience replay (IPER) strategy is adopted to increase the utilization of informative transitions and to enhance training stability. Simulations are conducted in static environments with different obstacle densities, and the proposed method is evaluated against representative reinforcement learning baselines and widely used classical planners. The results indicate that VT-ITD3 achieves reliable performance in terms of success rate, path efficiency, and trajectory smoothness across the tested scenarios. Further evaluations in multi-UAV navigation and dynamic obstacle settings, together with a PyBullet based physics simulation, demonstrate the applicability of the proposed method beyond static single-UAV navigation. Ablation studies further verify the complementary contributions of the dynamic virtual target generation mechanism and the improved replay strategy.