With the development of embodied intelligence technology and the accelerated industrialization of humanoid robots, insufficient data supply has become the core bottleneck for industry breakthroughs. The efficient processing system of multi-source real robot and simulation data opens up new paths for model training, while facing challenges such as cross-modal fusion and scene generalization. To address this, this paper proposes a full-lifecycle data management scheme to construct a systematic dataset system: starting from scenario engineering, covering robot body selection, hierarchical task planning, multi-modal collection, and full-chain quality inspection, integrating simulation and real robot data to form a virtual-real feedback loop. Validation on mainstream VLA models (π0) shows that the dataset produced by this scheme performs outstandingly in data diversity, scene coverage, and annotation accuracy, meeting the high-quality data requirements of embodied intelligence models and providing underlying support for improving model robustness and generalization.