Safe Path Planning of UAV Based on Deep Reinforcement Learning and Model Predictive Control

Path planning for unmanned aerial vehicles (UAVs) in complex and dynamic environments poses significant challenges in terms of safety. This paper proposes a safe path planning algorithm that integrates model predictive control (MPC) with soft actor-critic (SAC) reinforcement learning. The method uses MPC to perform real-time safety correction on actions generated by SAC, preserving exploratory benefits while ensuring compliance with dynamic and safety constraints. Simulation results demonstrate that SAC-MPC significantly reduces constraint violations in both static and dynamic obstacle scenarios. It achieves task success rates of 100% and 98%, respectively, confirming its effectiveness and robustness in complex environments.