<p>Parafoil systems are increasingly deployed in military and civilian missions that require high-precision trajectory tracking. However, conventional controllers based on linear strategies and limited decision information often fail to cope with task complexity, and existing reinforcement learning studies rarely consider the systematic construction of training task sets, thereby limiting generalization. To address these issues, this paper proposes a Benchmark Task Set Generation (BTSG) algorithm that builds high-quality task sets with diverse and balanced complexity for robust controller training and comprehensive evaluation, formulates the trajectory tracking problem as a Markov decision process with a 46-dimensional observation space and a dense reward structure to enrich decision-support information and improve learning efficiency, and develops a BTSG-SAC framework that integrates BTSG with Soft Actor-Critic (SAC) for controller training. Simulation results show that BTSG-SAC markedly outperforms PID and PPO controllers on the generated benchmark task set, reducing distance deviation by 50.7% and 33.2% under noise-free conditions; under sensor noise, BTSG-SAC attains a 98% success rate, whereas PPO reaches 77% and PID fails to achieve successful tracking. Overall, this paper extends the application of reinforcement learning to parafoil control and offers generalizable methodologies for future autonomous control research.</p>