Unmanned Aerial Vehicles (UAVs) operate in environments characterized by dynamic changes, structured uncertainties, and adversarial disturbances. These conditions require robust control strategies to ensure reliable performance. This study proposes a novel adversarial robust reinforcement learning framework that integrates adaptive adversarial training with self-supervised contrastive learning at the feature level. The adversarial network generates challenging scenarios by providing adversarial actions, which not only push the protagonist network to develop a generalized strategy but also serve as high-quality samples for contrastive learning. Integrating contrastive learning as an auxiliary task enhances feature diversity and reduces dependence on immediate rewards. This approach enables the policy network to capture the semantic structure of uncertainty and improve generalization to complex and unpredictable scenarios. It also helps maintain consistent and stable control actions even under severe adversarial conditions. Experimental evaluations of UAV tasks under environmental disturbances show that the proposed framework provides a reliable and adaptive solution for UAVs operating in challenging environments.