A Comparative Analysis of Reinforcement Learning Methods for UAV Autonomous Landing

Autonomous landing of Unmanned Aerial Vehicles (UAVs) is a sequential decision-making task in which an agent must autonomously select a sequence of actions to guide a vehicle safely onto a landing platform. This paper presents a comparative analysis of different reinforcement learning algorithms applied to the UAV landing problem, including Deep Q-Network (DQN), Soft Actor Critic (SAC) and Actor–Critic with Experience Replay (ACER). The landing task was formulated with a compact state representation and a discrete action state. A two-zone reward shaping scheme was developed to encourage horizontal convergence during approach and prioritized vertical descent in the landing region. All algorithms were trained and evaluated in simulation. The results show that SAC and ACER converge faster than DQN during training. In the testing results, DQN and SAC achieved 100% task success with an average landing deviation of 0.086\,m and 0.150\,m, respectively, while ACER reached 97% success with an average landing deviation of 0.141\,m. DQN had the smallest average landing error, SAC obtained the lowest average episode length and ACER exhibited the lowest crash termination during training. A comprehensive analysis of these results indicates a trade-off between precision and sample efficiency.