<p>Visual-inertial odometry (VIO) systems often struggle in visually challenging environments due to two common issues: ensuring reliable feature tracking and managing uncertainties that arise from integrating diverse observations. This paper presents a robust multi-camera VIO system that addresses these limitations through two key innovations. First, we develop a learning-based feature extraction module specifically designed for multi-camera configurations, which generates stable descriptors to validate optical flow correspondences and retain only high-quality features while reducing computational overhead. Second, we introduce an adaptive weighting scheme in the backend that employs weighting techniques across multiple cameras to quantify tracking uncertainties and assign appropriate weights to different viewpoints based on their reliability. The system seamlessly integrates an arbitrary number of cameras to enhance robustness in challenging scenarios. Extensive evaluation on datasets featuring visually degraded conditions demonstrates the effectiveness of our approach, achieving up to 90% reduction in ATE compared to state-of-the-art VIO methods.</p>