This paper introduces UniV2X, a novel end-to-end framework for cooperative autonomous driving that integrates key driving modules across diverse views into a unified network. The framework leverages V2X communication to combine data from both ego-vehicle and infrastructure sensors, enhancing perception, mapping, and occupancy prediction. A sparse-dense hybrid data transmission and fusion mechanism is proposed, offering three key advantages: 1) effective enhancement of agent perception, online mapping, and occupancy prediction; 2) transmission-friendly for practical communication conditions; and 3) reliable data fusion with interpretability. The framework is implemented and tested on the DAIR-V2X dataset, demonstrating significant improvements in planning performance and intermediate outputs. Experimental results show that UniV2X outperforms existing methods in collision rate, detection, and occupancy prediction. The framework also addresses challenges such as limited communication bandwidth, temporal misalignment, and data corruption, ensuring reliable and interpretable data transmission. The contributions include the first end-to-end framework for vehicle-infrastructure cooperative autonomous driving (VICAD), a sparse-dense hybrid transmission approach, and the reproduction of several benchmark methods. The framework is trained using imitation learning and self-supervised learning, ensuring effective and efficient planning. The results highlight the effectiveness of the sparse-dense hybrid data transmission in enhancing planning performance while maintaining communication efficiency and reliability.This paper introduces UniV2X, a novel end-to-end framework for cooperative autonomous driving that integrates key driving modules across diverse views into a unified network. The framework leverages V2X communication to combine data from both ego-vehicle and infrastructure sensors, enhancing perception, mapping, and occupancy prediction. A sparse-dense hybrid data transmission and fusion mechanism is proposed, offering three key advantages: 1) effective enhancement of agent perception, online mapping, and occupancy prediction; 2) transmission-friendly for practical communication conditions; and 3) reliable data fusion with interpretability. The framework is implemented and tested on the DAIR-V2X dataset, demonstrating significant improvements in planning performance and intermediate outputs. Experimental results show that UniV2X outperforms existing methods in collision rate, detection, and occupancy prediction. The framework also addresses challenges such as limited communication bandwidth, temporal misalignment, and data corruption, ensuring reliable and interpretable data transmission. The contributions include the first end-to-end framework for vehicle-infrastructure cooperative autonomous driving (VICAD), a sparse-dense hybrid transmission approach, and the reproduction of several benchmark methods. The framework is trained using imitation learning and self-supervised learning, ensuring effective and efficient planning. The results highlight the effectiveness of the sparse-dense hybrid data transmission in enhancing planning performance while maintaining communication efficiency and reliability.