Real-time trajectory generation for rapid pose maneuvering of fixed-wing unmanned aerial vehicles

In dynamic air combat environments, rapid pose maneuvering of fixed-wing unmanned aerial vehicles (UAVs) is crucial for establishing advantageous tactical postures. However, maneuvering trajectory generation for fixed-wing UAVs is a non-convex problem characterized by high-dimensional state and control spaces, making it a challenge to simultaneously ensure algorithmic computational efficiency and the dynamic feasibility of high-quality trajectories. To address this challenge, this paper proposes a learning framework based on Enhanced Proximal Policy Optimization (EHPPO). First, a rigorous constraint set is established based on the full-state dynamic model of fixed-wing UAVs, which ensures the dynamic feasibility of the trajectories. Second, a key-state-based piecewise reward gradient design is proposed, which facilitates the formation of accurate, rapid, and stable learning gradients, thereby ensuring trajectory quality. Third, by designing a group-normalization advantage function and an adaptive learning mechanism, EHPPO addresses the issue of poor learning consistency across target poses of varying difficulty and significantly enhances training stability. Furthermore, ablation experiments confirm the effectiveness of the reward design and the training strategy. Comparative experiments demonstrate that EHPPO outperforms both the Dubins-curve-based and MPPI methods in trajectory quality. Moreover, EHPPO improves computational efficiency by over two orders of magnitude compared to the Radau pseudo-spectral method, while maintaining a substantial advantage over MPPI. Unlike MINCO, our method empirically ensures the dynamic feasibility of high-quality maneuvering trajectories. Finally, the policy network is deployed in a C++ environment, achieving millisecond-level efficiency to realize the real-time generation of dynamically feasible, high-quality trajectories.