Hybrid MPC-RL Approach for Enhanced UAV Trajectory Tracking and Obstacle Avoidance

Model Predictive Control (MPC) has been widely applied in unmanned aerial vehicle (UAV) control. However, tuning the weight parameters in the objective function remains a major challenge. This paper proposes a UAV trajectory tracking and obstacle avoidance framework that integrates MPC with Reinforcement Learning (RL). By employing MPC as a function approximator within the RL framework, the proposed approach leverages MPC’s real-time constraint handling and disturbance rejection capabilities while utilizing RL’s global optimization to automatically tune MPC parameters, including cost function weights and model dynamics. Then, the classical Q-learning algorithm, driven by temporal difference, is employed to update the policy parameters. Experimental results demonstrate that the MPC-RL framework significantly improves trajectory tracking accuracy and obstacle avoidance robustness in both static and dynamic environments while reducing dependence on manual parameter adjustments.