Unmanned aerial vehicles (UAVs) autonomously control in narrow, visually complex tunnel environments require precise and robust obstacle avoidance strategies. Traditional reinforcement learning methods rely heavily on manual reward function design, often leading to inaccuracies due to human biases. Adversarial Inverse Reinforcement Learning (AIRL) presents a promising alternative, allowing UAVs to infer optimal reward functions directly from expert flight trajectories, thus eliminating manual task modeling. Nevertheless, AIRL frequently encounters training instability issues caused by adversarial structure, limiting the consistency and ability of learned behaviors and rewards. Thus we propose Bayesian Adversarial Inverse Reinforcement Learning (BAIRL), a novel variant of AIRL that integrates Bayesian neural networks to explicitly model uncertainty of reward function. BAIRL significantly reduces training fluctuations and converges faster to policies closely resembling expert trajectories. Furthermore, BAIRL enhances learning efficiency and adaptability when navigating complex environments without reward function modeling. Experiments conducted in a challenging visual-based tunnel scenario with continuous action spaces demonstrate that BAIRL achieves superior stability in narrow-space obstacle avoidance compared to traditional AIRL and imitation-learning methods, highlighting its substantial practical advantages for real-world UAV applications.