Learning Minimum-Time Flight for Fixed-Wing UAVs with High-Fidelity Dynamics

We address the minimum-time flight problem for fixed-wing UAVs, considering control constraints and fully exploiting UAV dynamics. Previous works have relied on simplified dynamics or path-planning methods that do not fully exploit the UAVs’ actuator capabilities. To this end, we apply deep reinforcement learning (DRL) with high-fidelity dynamics to train robust neural network controllers for minimum-time flight of fixed-wing UAVs. The results demonstrate that our approach is highly compatible with various DRL algorithms. Experiments conducted towards different target directions show that the method exhibits strong generalization capabilities and approaches time-optimal performance. In particular, the Proximal Policy Optimization (PPO)-based method exhibits robust convergence and stability in the trained neural network controller.