We present a novel visual odometry (VO) algorithm that achieves zero-shot generalization across diverse cameras and environments, addressing traditional limitations in VO algorithms associated with specific sensors and predefined settings. Our approach incorporates three main innovations. First, we introduce a language-based prior that infuses semantic information, enhancing robust feature extraction and enabling effective generalization to previously unseen domains. Second, we design a calibration-free, geometry-aware network structure capable of handling noise in estimated depth and camera parameters. Third, we demonstrate that our flexible architecture can leverage an unconstrained, semi-supervised training process that iteratively adapts to new scenes using unlabeled data, further boosting its ability to generalize across diverse scenarios. We focus on autonomous driving contexts and validate our approach extensively on three standard benchmarks—KITTI, nuScenes, and Argoverse 2—as well as a newly generated, high-fidelity synthetic dataset from Grand Theft Auto (GTA). Our work advances the boundaries of VO applicability, offering a versatile solution for real-world deployment at scale.