To boost UAV obstacle avoidance ability, this paper presents a UAV obstacle avoidance planning approach using an improved soft actor-critic (SAC) algorithm. It integrates course learning policy, which starts in a simple environment. Gradually increase the training difficulty, so that the UAV gradually improves the obstacle avoidance skills. A reward function based on the artificial potential field method is devised to guide the aircraft’s decision-making. A self-attention mechanism is introduced to better capture obstacle information. The input images are combined with a long short-term memory(LSTM)mechanism, allowing the aircraft to leverage historical data for optimized decision-making. Two experience buffer pools are set up to separate successful and failed flight experiences, and a priority experience replay mechanism based on temporal difference (TD) error is introduced to improve learning efficiency. Unreal Engine is used for simulation experiments, with real-time adjustable static and dynamic obstacles, and various weather conditions added to enhance algorithm robustness. Simulation results demonstrate a remarkable improvement in the obstacle avoidance ability of the improved SAC algorithm.